Journal oJ Glaciology, r ol, 42, Ao, 141, 1996 Avalanche forecasting-an expert systeDl. approach J GR G S CHW EI ZER A KD P AU L M, B, F OH:\" Eidgellossisches i ns lillll fill' Sdmee - llnd Lawinenforschung, CH-7260 /rJleissflllhjoclz /D avos, Swi lzerland ABSTRACT, Al ala nc h e fo recas ting fo r a gi,'e n region JS still a diffi c ult tas k im 'o h-ing g r ea t res ponsibility , An y tool s as istin g th e ex p e rt in th e d ec isio n-m a king process a r c we lcome, H oweve r , a n effi cient a nd success ful to ol sho uld mee t th e n eed s of th e forecas te r, vVith thi s in mind , (wo m o d e ls, )ve re d evelop e d usin g a co mm e rcia ll y al'ail a bl e soft wa re: CYBER TEK -COGENSY S I CIf , a jud g m e nt processor fo r indu cti ve d ec ision-m a kin g - a prin c ip a ll y d a ta -b ased ex p ert sys te m, Us in g wea th er, sn ow a nd snow- cove r d a ta as inpu t p a r a m eters, th e m o d els eva lu a te fo r a region th e d eg ree of al'a la nch e h aza rd , th e as p ec t a nd altitud e o f th e mos t d a n gero us slopes, Th e o utput res ult is b ased o n th e snow- cO\'e r sta bility , Th e new model s w er e d eve loped a nd have bee n tes ted in th e Da l'os r eg io n (Swi ss Alps) fo r seve ra l yea r s , T o ra te th e m o del s, th eir o utpu t is c omp a red to th e a p os teri ori I'e rifi ed h aza rd, Th e fir st model is p urel y d ata- based, Comp a red to o th e r sta ti sti cal m o del s, th e diffe r e n ces a rc: m o r e input inform a ti o n a bou t th e sn o w cove r fro m sn o w profil es and Rutsc hbl oc k tes ts, th e spec ifi c m e th od to searc h for simil a r situ a ti o ns, th e co n cise o utput res ult a nd th e kn ow led ge b ase th a t includ es th e I'e rifi ed d eg ree o f ava la nche h aza rd, Th e perfo rm a n ce is a bout 60 %, Th e seco nd , m o r e-refin ed m o d e l, is both d a ta - a nd rul e- based, It tries to mod el th e d ec ision-m a kin g process of a prag m a ti c expe rt a nd has a perform a n ce o f a bout 70 0/0 , w hi ch is co mp a ra bl e to th e acc uracy of th e publi c ,\'a rnin g , 1. INTRODUCTION AI'a la nc h e for ecas tin g, in our co ntex t , m ean s th e da il y assess m e nt o f th e a l'a la nc h e h aza rd for a g iven region, i, e, fo recas tin g a t th e meso -sca le (M eClun g a nd Sc hae rer, 1993 ) , Th e res ultin g ava la n c h e wa rnin gs a nd reco mm en- d a ti o ns [o r th e publi c sh o uld d esc rib e th e a l'a la nche situ a tio n , i, e, gi,'e info rm a ti o n a bout th e pl ace, th e tim e a nd th e pro bability o f r e lease for a sp ec ifi c type of a l'a la nch e (sla b or slufT, la r ge o r sma ll , we t o r dry), Th e mos t co nl' e ni ent way to ha ndl e thi s so rt o f inform a ti on is to summ a ri ze it as a d eg r ee of al'a lanch e h aza rd, Si nce 1985, in Switze rl a nd , th e d eg ree of h aza rd has bee n d efin ed in d esce ndin g ord e r b y th e release pro b a bili ty , th e a rea l ex te n t of th e in st a biliti es a nd th e size of th e a l'a la n ch es (F ohn , 1985 ) , Th e sca le is ge n e r a ll y based on th e sn ow-cOl'e r sta bilit y, I t co pies th e d evelopm ent o r stepping o f th e mos t typi ca l al'a la neh e situ a ti ons a nd hence is no t lin ea r. I 'h e int ensit y o f a n ava la nche situ a ti o n in c reases stro ng ly fr om one d eg r ee to a no th er , m ay be el 'e n ex p onenti a ll y , Co nsequ entl y, th e fr equ ency of th e d eg rees of haza rd d ec reases acco rdin g ly with in cr easin g d egr ee of haza rd (Fi g, I ), An y ex p ert sys tem should p ro fit from thi s co n ce pt th a t was a d o pt ed in 1993 b y th e w o rkin g gro up o f th e Eu ro p ea n al'a la nche- wa rnin g se n 'ices , In thi s stud y, seve n d eg r ees o f ava la nc h e haza rd a re u sed acco rdin g to t h e stru cture d efin ed in 1985; d eta il s o f th e Swi ss haza rd sca le hal'e a lso bee n giv e n b y Nl cC lun g a nd Sc hae rer ( 1993 ) , Sin ce dr y-sla b ava la n ch es represe nt th e m os t impo r- ta nt threa t fo r ski ers a nd bac k- co untr y tra l'cll ers, \\'e 3 18 degree of haza rd Fig , 1, R ela tive .frequeJlc} oJ tlt e verified degree oJ haz ard ill Ihe D avos region : len winter seasolls including 1512d are considered, Lighl columns ( lift) Io r the old Swiss sel'fll -degree scale, dark [olul11ns ( right) and valuesJor the Ilew European Jive-degree scale , foc used o n th e hazard of dr y-sla b ava la nc h es, During sprin g tim e, ,,'e t-snow al'a la nc h es a re pa rtl y co n sid ered: th e d ail y in cr ease of th e haza rd du e to wa rmin g durin g th e d ay is no t ta ken into acco unt , LaC ha pcll e ( 198 0 ) d esc rib ed th e tec hni q u e fo r assessin g th e ava la nche h aza rd: wea th er , sn o w a nd snow- cove r d a ta o bse rved d a il y a nd meas ured a t se l'er a l loca ti ons r eprese nta ti,'e for a g ive n area a re eva lu a ted by hum a n ex p e rts usin g th eir kn o wl edge a nd lo ng -term ex peri ence combin ed with indi vidu al intuiti o n, Sin ce th en th e procedure has not c h a nged mu ch, Th e core is still form ed b y th e cl assica l process of sy no psis suppl e- Downloaded from https://www.cambridge.org/core. 06 Apr 2021 at 02:05:53, subject to the Cambridge Core terms of use. https://www.cambridge.org/core Sclzweizer al/d FiJllIl : .4 volallclte forecasti ng - an e.\/)erl s),slell1 a/Jproaclt T : verif ica tion : sup po rting t ools: st ati sti cal determini stic expert systems Fig , 2 , The clo:,sical cOll vellliol1al melhod sll/J/)Iel/lenled willt differenl sllpjJorting lools 10 forecasl tlte allolanc/ze /1Ci::ard on l/ze regiollal scale, me nt ed n owa d ays by diffe re nt so rt s o f s u p po rtin g too ls (Fig,2 ) , H owel'e r , th e d e m a nd s are stead il y in c reasin g : m ore fi'Cq u e n t a nd m ore d e t a i led in for m a ti o n is d es ir ed, T o uri sm is sti ll d eve lo pin g a nd , as in m a ny o th e r mou n ta in o u s reg io ns, th e A lps ha l'C b ecome o nc of th e fa l'o urit e p layg ro und s in Euro pe, S ki lO ur ing is wid e r ra ng in g a n d mo re p o pul a r th a n it was prel' io usl y , Neve rth eless, th e num b er 0 [' a\'a la ne h e I'ic tim s has n o t acco rdin g ly in c reased (Fi g, 3 ), whi c h m ay be du e to th e be ll er edu ca ti o n a nd aware n ess of th e ski e rs, pres um a bl y a lso d u e lO th e stab ili zin g e ITec t of m o r e fr eq uent skiin g a ft e r eac h s noll'fa ll per io d o n po pul a r slo p es a nd ho pefu ll y du e to b e tter \\'a rni ng, In th e fo recast pr ocesses nO\I'adays, a lo t 0 [' electro ni c too ls a rc i Jl\'o lved: ac q UI S111 0 1l , tra nsfe r an d represe n t- atio n o f t h e da ta, da ta b ase, snow -co l'e r si m ul atio n , 1111 free terrain ~ traffic o buildings ~ 60 r-------------------------------~ ~ '5 :§ 40 Q) .J:. g 20 o '5 > o o 65 70 75 80 85 90 95 year Fig , 3, Amlanche falalilies ill Ihe SwiH : II/)s 1964 6510 J99-1- 95, The falalilies are appoillled 10 Iltree calegories describing T('herl' Ihe {wa/al/che accidenls hajJjJelled: in blli/dillgs, 011 COlIll/IUllicalion lilles (illcllldillg cOlllrolled ski runs) and in Ihefree lermin ( back (0111111],), The 30)'ear average is aboul 28 falalities, indicaled b)' l/ie Ihill broken line , T he solid bold lille shows 0 5 ) 'ear IIlOl'illg average , Percentage Illlmbers give Ihe /)I'o/)orliol1 of falalities ill Ihe free terrain ( 10 ) )eal' average) showing deor~y Ihe increasing number I!/jalaii lies in Ihe free lerrain, whereas lite lolal IIlImber Irend is slalional)'. n u m erica l wea th e r fo recas t , d ec ision -ma kin g too ls a nd ex pe rt syste m s, a nd info r ma ti o n di stributi o n , EI'e n so, th e d ata t ha t a re m eas ured a re n o t th e mos t re leva n t o n es, \Yh a t is rea ll y n eed ed is th e s tre n g t h (co mpress ive, te nsile a nd shear ) of indi vidu a l sn ow laye rs, the so -ca ll ed 10 11'- e n t ro p y d a l a ( L aC h ape ll e, 1980 ) o r cl ass I d a ta (M cC lun g a nd Sc hae rer , 1993 ) , Th e a l'a il a b le d a ta a rc jus t mo re o r less app ro pri a te to p a r a m e te ri ze th e re!el'a nt p rocesses , T h e int erpre ta ti o n o f th ese d ata is t h e m os t d e li ca te tas k a nd hence m a n y su ppo rtin g too ls h ave bee n d el 'C lo ped fo r hum a n expe rts, Because th e 3\'a lanc he h aza rd ca nn o t (ye t?) be ca lcu la ted full y in a st ric t se nse ( by a lgo rithm s) , thi s is a fi e ld fo r hu ma n ex p e rts a nd co rres po ndin g ly for ex pe rt sys te m s, 2. PRESENT APPROACHES Th e a p p roac h d esc rib ed b y L aC h a pe ll e ( 1980 ), th e c lassica l m e th od , sti ll fo rm s th e b as is of th e d ec isio n- ma kin g proced ure of mos t ava la n c h e-fo recast sen' ices , U p till now, n o n e o f th e supp o rtin g tools have b ee n re li a b le e n o ug h to s ubs titute fo r th e hum a n expe rt a nd will proba bl y n e l'Cr bc, But m ay th ey beco m e a n obj ec til 'e p a rtn er fo r " di sc uss io n "? A ge n e ra l O\'crl'iew 0 [' d iffe re nt me th ods has b ee n g il'en by F b hn a nd o th e rs ( 1977 ) , Buse r a nd o th e rs ( 1985 ) a nd m o re rece n l ly by l\l cC lun g a nd Sc hae rer ( 1993 ) , Ln th e fo ll ow in g we foc us o n fo recas tin g mod els a nd too ls, Statistical approache s Th e mos t po pu la r sta ti sti ca l m e th od s a re t he di sc rimin a nt ana lys is a nd t h e nea res t n e ig h bo urs (:\I cC lung a nd Sc hae rer, 1993 ) , Already, in th c 1970s th e firs t s tudi es usin g d iscr imin a nt a n a lys is h ad b ecn pe rfo rm ed to find th e re le l'a nt pa ra mete rs fo r <1 I'a la nche fo recas tin g (e,g , P e d a, 1970; J udso n a nd Eri c kso n , 1973; A n nst ro ng a nd o th ers, 197+: Bo is a nd o th e rs, 1975; Sa l way, 1976 ), Snow a n d weat he r d a ta a rc us u a ll y use d toge th e r w it h obse n 'a ti ons of al'a la nche ac ti l' it y, ~ew sno ll' dep th , tem pe ra lUre a nd win d spee d , to m e nti o n so m e, h a l'e pro l'e d to be i m po rta n t. Th e resu lts ha I'C co n fi rmed th e ex pe ri e nce o f th e al'a la nc he expe n s, But , as n one of t he pa r ame ters used is d irec tl y re la ted to t he p rocess 0 [' a l'a la nc he fo r ma ti o n, it h as not bee n p oss ib le to el'a lu a te t h e al'a la nc he h aza rd, Th e d a ta uscd , th e usu a l o bse n 'ed a nd m eas ured pa ra me te rs, a r e a ll ind ex va lu es, (Th e d ata used a r e t hose th a t a re a l'a il ab le a nd no t th ose th at a rc m os t rcle l'a nt to th e process o f ava la nche fo rm a ti o n ,) Th ey a re in s tru c til'e to a n ex pe rt a nd may g ilT th e co r rec t hints to th e key p rocesses, su c h as se tt lem e nt. H Ovl'e\'e r, th e res u lts o f'th ese sta ti stica l studi es hal'e im p roved th e und ersta ndin g a nd h a l'e hclJJ ed lO s tru c ture know le dge a nd fin a ll y to d el'C lo p rule-b ased sys te m s, Additi o n a ll y, as th e sta tis ti ca l m od els n eed lo ng- te rm da ta, ma n y I'a lu a bl e o bse n' a ti o n s hal'e b ec n init ia ted, T h e acc umu la ted d a ta base m ay n olV be used to im p ro l 'e th e me m o r y o f' th e expe rt. Op era ti o n a l sys tems based o n th e sta ti sti ca l a pproac h, a nd usin g a lo n g- term d a ta b ase, h ave bee n d eve lo ped in sCI'C ra l co untri es an d a rc n o \\' w id e ly used (Bu se r a nd o th ers, 1987; N ava rre a nd o th e rs, 1987: M cC lun g, 1994; 3 19 Downloaded from https://www.cambridge.org/core. 06 Apr 2021 at 02:05:53, subject to the Cambridge Core terms of use. https://www.cambridge.org/core J ournal of Glaciolog)I M cC lu ng and Tweedy, 1994; }. I crind o l, 1995 ) bo th for loca l and for reg iona l ava la nche fo recas tin g . Except fo r rh e sys te m d eve loped b y M cClun g a nd Tweedy (1994) , whi ch is a co mbin at ion of the tw o stat isti ca l m et hod s, all o f'th em use th e nea rest-neig h bour meth od. It is ge nera ll y assumed th a t simil ar snow a nd weather co ndition s should lead to simil a r ava la nche situ a tion s, i.e. that obser\'ed a\'a la nches of's imil ar past da ys sho uld be rep rese ntat i\'e ofth c prese nt- day situ a ti o n. A geo me tri ca l di stan ce in th e input p a r a meter space is used in sea rch i ng for simil a r si tu a - tion s. Th e Euel idi an di stance betwee n th e ac tual-d ay data a nd the surroundin g pas t-da y data is in so me mod els ca lculat ed direc tl y in th e input para m e ter space (Bu se r and ot h ers . 1987 ); in ot her mod els, the input d ata a rc first t ra nsform ed in the space of th e prin cipa l co mpon ents that w as determin ed b y th e sta ti sti ca l a nal ys is o f the d a ta base (lV[ crindol , 1995 ) and then th e M a ha la nobi s di sta nce is used (M cClun g and Tw eed y, 1994 ) . In so me models, the input para mete rs a re weig hted accOl-din g to the ge n eral ex p erien ce o f the expert fo recas ter (Bu ser a nd o th ers, 1987 ) and th e we ig hts may eve n vary accordin g to th e ge nera l wea ther type (Bolognesi , 1993b ) . The output is ge nerall y bas ed on th e obse rved avalan ches of' th e ten or 30 neares t n eig hb o urin g d ays . Thi s info rm a tion has to be e\'a lu ated by th e for ecas ter. Th e simplest way to summariz e thi s so rt of inform a ti on is to just se para te betll'een avalanche days a nd non -avalanche da)'s (Obled a nd Good , 1980 ). In add iti o n to th e li st of th e 30 nea rest neig hb ours , M cC lun g and Tw eed y ( 1994 ) predi cted th e p roba bility o f a\'a la nching by using di sc rimin a nt ana lysis. For th e fo recastin g of a \'a lan ches at K oo te nay Pass (B.C., Canada ) th is pro ba bility is co m- bin ed , usin g Baysian statisti cs , with the ex pert 's op ini on , m a d e a pri ori , to tak e into account a ddition allo\\'-entropy d a ta (e.g . th e sn ow-cO\'er situ ation ) (W e ir a nd }'[CC lun g, 1994 ) . In oth er m od els. a number is g iven as o utput acco rding to th e class ifi ca tion o f o bse rv ed avala nch es used in th e country (Gu yo marc' h a nd M e rindol , 1994 ) o r a n a va lanc he in dex is ca lcul a ted (N avarre a nd o th ers, 1987 ). Th ese types of o utput are difIi cult to r e late to the ac tual h aza rd in a give n reg ion. H ence, it is difIicul t to assess th e r ea l quality o f th ese fore cas t models. The y ce rtainl y improve th e refle c ti ons of un ex pe ri enc ed forecaste rs a nd m ay influ ence ex p eri enced forecasters but rarely may the y be ca lled a dec isi\'e help in dete rminin g the d eg ree of h aza rd in a given region. DeterIIlinis tic approaches Th e a im of the d e te rmini stic a pproach is to simul a te th e ava la nch e rel ease. \\' hereas th e m os t rel eva nt snow -cover pro cesses may be mod ell ed (Bru n a nd ot hers, 1989, 1992 ) and so m e a ttempts ha \'e a lso been m a d e to model th e ava lanc he fo rm a ti o n (G ubl er a nd Bader, 1989 ), it is still a lmos t imp ossib le to simulate t he range of the num erou s avala nche-fo rm a tion processes o n a m ountain slope - not to sp eak of a lI'ho le a rea. A poss ibl e way out is to use d iffer ent me thod s, fo r exa mpl e, to d eve lo p a n ex p ert sys tem w hi ch a n a lyzes the simulated snow-cove r st r a t- ig raph y (Giraud, 199 1). Fohn a nd H aec hl er ( 1978 ) d e\'clo ped a determ ini sti c- sta ti sti ca l m od el whi ch re lates t h e snow acc umul a ti o n by snowfa ll , wind a nd se ttl e m ent to th e a\'a la nche ac tivit y. Th e mod el is a ppropri a te to d esc ribe a\'a lanc he situ at ions in pe riods of heavy snowfa ll. 32 0 E x pert systeIIls Exp er t sys tem s· represe nt the id ea of simul at ing th e d ec isio n-m a kin g process of a n expe rt . Mos t o f them a r e symboli c co mputin g sys t e m s, i.e. usin g rul es w hi ch we r e formulat ed ex pli citl y by hum a n expe rt s, e.g . }' IEPR A (G ira u d , 199 1) a nd A V ALOG (Bol og n es i, 1993a ) . The Fr e n c h sys tem l\lEPRA a nal yzes the sn ow -cove r stra tig r a ph y; th e sno w profi les a re simul ated b y th e snow- cover mod el C RO CUS (Brun a nd others , 1989 ) in p a r a ll e l w ith m e t eo r o logica l data prO\' id ed by SAF R A J (Dura nd and o thers, 1993 ) , a m ode l fo r o ptim a l int e rpo latio n of m eteo ro logical d a ta. AV A - LOG , assessin g th e ava la n ch e hazard slope-b y-slope, is a n ass ist in g too l fo r th e e ffi cient art ifi cia l release by ex plosives in the res tric t ed a rea ofa ski reso rt. R ece ntl y, a hybrid exp er t sys te m h as bepn d eve lop ed usin g a neura l n e twork a nd rules ext racted from th e data base with n e ura l netwo rk techniqu es (Schweize r a nd o th e rs, 1994a, b ) . Bolog nesi ( 1993b ) h as d e \' e loped ano th er h y brid sys tem ca ll ed NX-LOG b y co m bi ning the statis ti cal mod el NXD (Bu ser and ot h e r s, 1987 ) and th e rul e - based syste m AVALOG. A furth er rece nt d evelopm ent is a rul e-b ased ex pe rt syste m to interpre t data fr o m snoll' profi les with respect to snow stabilit y (M cClun g , 1995 ). 3. A NEW APPROACH WITH THE CYBERTEK. COGENSYS™ JUDGMENT PROCESSOR In 1989 , we bega n a n ew ap proac h with th e id ea of buildin g a sys tem for region a l a\'a la nche fo r ecas tin g simil a r to th e sta ti stic a l ones bu t wi th a d i fferen t method of sea rchin g for simi lar situations and with optimi zed inpu t a nd output pa ra me ters, ca ll ed DAVOS. \V e tri ed to inclu de so m e of the rel eva nt ph ys ica l processes, i.e. elab o ra ted in put parameters a nd to give as a res ult direc tly what th e ava la nc he forecaster wo uld like to have: th e degree of haza rd (Schweizer and o th ers, unpublished ) . In 199 1, we worked o u t a co mplet ely new app roac h , more process -oriented a nd part ly r ul e-based , wh ich tri ed to mod e l th e reaso nin g of th e ava lanc h e for ecas ter, ca ll ed l\ IOD UL. Both models a re ba sed on softw a re for indu ctiv e d ec ision -m a kin g : CYBE R TEK-COGENSYST~ I jud g - me nt pro cesso r (ve rsion 19 ) , which is prim aril y used in th e fin a n ce a nd in suran ce wor ld . The IIlethod: the judgIIlent proc e ssor The C YB E RTEK- COGENSYSD I jud g m ent processor is a co mm e rc ia ll y a\'a ilabl e software fo r indu cti\'e a uto - matic d ec ision-m a kin g . Sin ce we had no access to th e so urce cod e, we did not exact ly know what the sys tem does A traditi o na l ex pert syste m co nsists o Cfi ve eleme nts: th e dial og ue co mpo nent, the prob lem- so lvin g co mpon ent , th e kn ow led ge base, th e ex plan ato r y compo nent and th e co mpon ent for in c remental lea rnin g (D U DEN Informatik , 1988 ) . Downloaded from https://www.cambridge.org/core. 06 Apr 2021 at 02:05:53, subject to the Cambridge Core terms of use. https://www.cambridge.org/core SC/z11'ei::.er and FiJ/m : A1'OIall che Jorecastillg - an eljJerl 5,J'stelll ap/noarl! (h oweve r, it works ) . So , th e a lgo rithm ca nn o t be g iven in all detail, but w e sha ll tr y to o utlin e th e ge n era l id ea b e low . So far , th e co re of the sys tem m ay be consirlered a " black box" . The judg m ent pro cessor is based on th e fact th a t prag m atic ex p e rts d ec id e, u sin g th eir ex perience a nd in tui tion rather than ex plici t rul es . The more co mpl ex a problem, th e less stru ct ured is th e kn ow ledge . Experts are u s u a ll y a ble to decide correc tl y and fas t in a real situ atio n . H owe \'e r, th ey a r e usuall y h a rdl y a ble to ex pl a in their d ec isions co mpletely following exact rules. Th e e xpert 's approac h is to c ho ose the rcl e\'a nt d a ta (which m ay diner s ubs tantia ll y from o ne situ at ion lo a no th er) , to elas. if I' a nd to analyze th e d a ta a nd fin a ll y to reac h a co nclu sio n. Buildin g up a mode l im'o h- es the fo ll ow in g s teps: 1. A so -ca ll ed judgment jJ1'oblem co n sis ts of a li st of question s a nd th e logic r eq uired to a rriv e a t ajudgme nt - that is, to reac h a co nclu sion or m a ke a d ec isio n - based o n the a nswe rs to those ques tion s. By specifying th e questi o ns - to be a ns\\' ered by yes/ no, mu ltipl e c h oice or num er ica l responses - th e m e nt o r, th e expe rt buildin g up the sys tem , d efin es th e d a ta needed to reac h a spec ific d ec ision a nd th e c rit eria that a re used to ca tego ri ze or e \'a lu a te th e data . Th a t m e an s fo r num eri ca l ques ti o ns th a t th e poss ibl e answe rs mu s t be g rouped into so -ca lled logical rallges (up to fiv e ranges), so th a t th e sys tem ca n lea rn h o\\' th e respon se is norm a lly categori zed. Num e ri ca l qu es ti o ns ca n ta ke the form of calcu lation s in cludin g co nditi ons. 2. On ce a problem h as b ee n d e fin ed, th e mentor " teac hes" the jud g ment processo r by ent erin g exam - pl es (rea l or r ea li sti c d a ta ) and inte rpre tin g the situ at io ns rep rese nted b\' th ose exa mpl es. By obse n 'in g the relatio nship be t\\'ee n th e d a ta an d th e mento r' s d ec isions . th e jud g ment pro cessor build s a logica l m ode l th a t a ll ows it to e l11ul a tC' the m e nto r' s dec isio n s. Th e more co mpl ex th e pro bl em the m o re situ at io n s arc needed. H o we\'e r, as usua l in case-b ased reasonin g sys tem s, th e pe rfo rman ce of th e sys te m in creases fa s t a t th e beg innin g with increasing numb er of silU a ri o ns, reac hes a plateau an d fin a ll y ma y eH' n decrease ( Fi g. 4 ). 3 . Th e judgm e nt processo r calcu la tes the so-ca ll ed logiral im/Jortall l'e of eac h qu es tion based o n th e obse rnl ti o n of th e ment or\ de c ision. Th e log ica l importan ce is a m eas ure ofho\\' a parti c ul a r ques t io n co ntributes to th e logica l model as a \\' hole, base d o n h O\\' ma ny situati o ns within the kn o w led ge base wou ld b eco me indi stin gu ish- a ble if th a t qu es ti o n were to be re m o \·ed. Base d o n th e log ica l impo rt a n ce, gi\'e n as a number fi'o m 1 ... 100 , t h e qu es ti o ns are class ifi ed as so - ca ll ed II/ajor o r min or question s. Th e log ica l im po rt a n ce is co n tin uou sl y upd a ted, so th e sys tem can lea rn in c reme ntall y . After suffi c ie nt tra inin g of th e m o d e l by th e m e nto r, th e m od el perform s th e following s te ps to r eac h a co nclu sio n for a n e w situati o n e nte r ed : 1. If a new situ a ti on is enco unt e re d , th e sys te m tri es to g ive a proposi ti o n fo r the poss i bl e d ec isio n o n rh e basis of rhe past kn ow n situati o n s and o n what is learned a bo ut th e d ec isio n log ic; p a rti c ul ar ly, the class ifi cation into m aj o r a nd min o r qu es ti o n s b ased o n th e pre\'ious ly ca lcu la ted log ica l imp o rt a n ce is O.B IJ) IJ) Q) 0.6 r:::: h 0 -(.) ~ 0.4 0 (.) 0.2 500 1000 1500 number of situations in knowledge base F ig. -1. A {),pical exam/Jle oJ th e pelformanre oJ a case- based reasoning 5,ystem with increasing knowledge base Ol'e r time: th e DA l'OS4 model . use d. A new situ a tion is simil a r to a kn ow n p ast situation from th e knowled ge b ase, if a ce rtain numb e r of the a n swe rs (usua ll y 80 % ) to the m ajor qu es ti o ns arc \\' ithin th e sa me log ica l ra nge . That me ans th at primaril y o nl y the m aj or qu es tion s a r e co n sid ered fo r sea rc hin g for simil a l- s itu at ions . Two pa s t situati o ns th a t arc bo th s imi lar to a n ew s itu a ti o n ma y h ence co in cid e in diffe re nt qu es ti o n s, e .g. if a prob le m co nsists of fi\ 'e major qu es ti ons, four a n s\\'C rs ha \'C to b e in th e sa m e ran ge, and he nce fi\'e diffe re nt poss ibiliti es ex ist 10 r a s imi la r situ at io n , in add iti o n to th e case wh en a pre \'i o us situ at ion is found th a t co in c id es in a ll an swc rs to th e m aj or ques ti o ns. So simi la r situati o ns n ee d not co n se q- u e ntl y be nea r , in a geo m e tri ca l se nse in th e parameters space , to the new situ a ti o n. T I. Th e pro posed d ec is io n is d e ri\ 'C d fr o m the similar s i tu a ti o ns fo und u s i ng the so -call ed (lssertioll !el'el of th e difTerent sim il a r situ a ti o ns. All qu es ti o ns a nd th e ir log ica l imp o rt a n ce arc conside r ed to d e te rmin e th e a sse rt ion leve l. rt is a numb e r (o n a sca le of 1- 100 ) that ren ec ts h o\\' close ly the c urren t situation co mp a res to ex is tin g situ a ti o ns ill th e knowl edge ba se . Th e close r th e a sse rti on le\'e l is to 100, the more simil a r th is examp le is to prev ious ly encountered s itu a ti o ns. Th e less th e answers ag r ee. th e sm a ll e r is th e a sse rti o n le \"e l, i.e. for eac h ans\\'e r th at does not ag r ee, a ce rta in a m o unt is subtr acted ri'om 100 , depending o n th e number o f qu es ti o ns a nd th e log ica l imp o rt ance of th e qu es ti on; in the case of p e rfec t agree m e nt , a so -ca ll ed full match, th e asse rti o n le\'e l is h e n ce eq ual to 100. rn . Th e qua lit y o f the pro posed deci s ion, ba sed o n t h e s imilar situ a tio ns fo und , is d esc rib ed b y th e so -ca ll ed cO lifidenre lerel, a n ind ica to r of how cer tain th e syste m is that its inte rpretat ion is approp ri a te to the curre nt situ a ti o n: a n ex cl a mati o n mar k ( I ) for \ 'ery confid e nt , a p e ri od (.) fo r I'easo nab ly co nfid e nt or a qu es ti o n mark (?) for not co nfid e nt. A low !c\'cl of co nfid e n ce s u gges ts th a t th e re a r e fe \\' situ a ti o n s th a t th e sys tem co n s id e rs to b e logica ll y simil a r , o r th a t th ose s itu a t io ns that a r c simil ar have co nOi c ting int e r - pre ta ti o ns. Additiona ll y, th e simil a r situ at ions th a t a r e u sed to d e ri\ 'C th e deci sio n with th e acco rdin g asse ni o n le\'e l a r e a lso gi\'e n as a n ex p la n a ti o n . J f th e sys te m is not ab le to lind a dec isio n o n th e bas is o f th e 32 1 Downloaded from https://www.cambridge.org/core. 06 Apr 2021 at 02:05:53, subject to the Cambridge Core terms of use. https://www.cambridge.org/core J ouma l ~! Glaciolog), prese nt kn ow led ge b ase it gives th e r es ult " not poss ible to m a ke a n int e rpre ta ti on", in t h e fo ll owing simpl y ca lled " no inr e rpre t a ti on" . In Fig ure 5 th e different s te ps to reae h a co nclu sion (d esc rib ed a b O\·e ) a re summ a ri zed in stro n g ly simplifi ed g ra phi ca l fo rm. In T a ble I a n exa mpl e o f th e sys tem output is gi, ·e n. Sin ce the sear ch fo r simil a r situ a ti ons fo rm s th e co re of th e m Clh od , it m ay be call ed , in the broa d es t se nse, a n ea res t-n e ig hb o ur met h od . H o ,,·e,·e r , th e m e tri c for sea rchin g fo r simil a r situ a ti o n s differs substa nti a ll y fr om th e co mmonl y used dista n ce m eas ure, e.g . th e Eu cl idi a n di sta n ce . Th e cat ego ri za ti o n of th e input d a ta, th e cl assifi ca tion int o maj or a nd min or qu es ti o n s a nd th e m e tric to sea rc h simi lar situ a ti o ns a re all no n -li n ea r. Brieny summ a ri zed , th e sys tem we ig h s a nd cl ass ifi es th e ca tegor- ized d a ta, sea rc hes fo r simil a r situ a ti ons stro ng ly using th e cl assifi ca ti o n a nd ca tegori za ti o n , d eri ves a res ult fr om th e simil a r situ at io ns, desc ribes th e qu a lity of th e r es ult a nd fi nall y li sts th e simil a r situ a ti o n s used for d e ri,·ing th e res ult 32 2 + and x: past situations with according output (+ o r x) 0 : new situation , output unknown similar situations (c) toget h er ,,· ith th e pert in ent s im il ar ity meas ur e. T he a dva ntage of th e m e th od is th e st ro ng co nce ntra ti o n on th e qu es ti ons th at a re co nsid ered im p o rta nt. Th e fac t th a t th e m a jorit y of th e a nswers to th e m aj or qu es ti o ns h ave to b e in th e sa me log ica l ra nge m a kes th e logica l imp o rta nce o[ a q u es ti on , co mp a ra ble to th e we ig ht used in a d iffe re nt sys tem , to a d ecisi ve [acLO r, in co ntras t to simil a r sys te ms. So differe nt ,·e rsio ns of a mod e l, as we will use, w i th th e sa m e qu es ti on s but with diffe r e nt logica l imp o rt a nce (ca lcul a ted by th e sys tem, no t g ive n arb itra ril y), leadin g to a differen t p a rtiti on of maj o r a nd m in or qu es ti o n s, ,\·ill find vc ry differ ent simil a r situ at io n s. The application: the avalanche hazard In o ur case th e judg m ent pro bl e m is th e avala n c h e h aza rd a nd th e qu es ti ons will b e ca lled input p a l-a m e te rs a nd a re, fo r exa m pie, th e 3 d s u m of n ew snow d e p t h or th e air temp e r at ure . Th e a n swe r s a re th e ,·a lu es o f th e input p a ra m e te r s in a rea l situ a ti o n , e .g . 15 c m o r - 5°C. A rea l situ a ti o n is he nce d esc rib e d by the se t o f in p ut Fig. 5. CrB E RTEK- COG E KS r ST. 1f J udgment P ro - cessor : the different ste/JS to reach a conclusion are given (strongly sil7l/JLified) Jor a problem with on0' three injJut parameters (X/, ){2, .1.3 ) and one output /Jarameter ("x" or "+") . (a) InjJu t /Jarwneter space . (b) Categorization, leading 10 a cube (x/ , \2, 1"3) of 125 identical boxes . (c) Reduction to the major /Jawmeters (x /, \"3), i.e. projection to this jJLane. Selecting simiLar situations (all situations in the shaded squares) based 0 11 the Jollowing similarilJ' conditio,, : similar situations are all past situations with either x/or X3 ill the same catego l)' as fhe lIew situalion . I II the sha ded sqllares are often several similar /Jas t situa tiolls; t/tese si tualions that may have different output differ ill the third (minor) parameter (X2) . Based on the similar siluations, referring to the Logical il7ljJortallce and the minor parameter . the system /Jroposes the result; ill the above case, the /Jroposed ou tput would be "+ ", with a confidence Level cif ")", i.e. not confident, since there are too mClll)' simiLar situations with different output. Downloaded from https://www.cambridge.org/core. 06 Apr 2021 at 02:05:53, subject to the Cambridge Core terms of use. https://www.cambridge.org/core Schwei;:er alld F iJlzn: Al'ala l/ che jorecastillg - all eX/Hrt s),slem ap/Jroach T able I" Exa mple oj screell oUlpu l: A"lODeL model, sub-problem Rel ease proba bilit y n ew snow " !l J anuary , 1995 (s ee ledjor eX/Jlallaliol/ ojlerllls and Figure 8) " T he sub-/JToblem has six ill/Jul !)(/)"{//nelers; the S) 's lelll /Jroposed the oll l/JUI la rgejor th e release probabiLiry , bu t it is /l ol SlITe at aLL, so it /JII IS a question mark" As explanation,jolll" similar situalions are gIVell U se r ID : Pro bl e m cod e : SLF RPROB_.\ S 316 Situ a ti o n N o: " \ ro" DATA 01eslion i\ ew sn o \\" d e pth (5 , 15 , 3 0 , 50 cm ) 3 d sum o f n ew sn ow d e pth ( 10 , 30 , 60 , 100 cm ) D a t e : Pro bl e m n a m e: Appli ca n t cod e: Answer 81 cm 93 1] ] allll(//Y 1995 09: 14 Release probabiLi~)1 new SIlO W 1/0195 (j L/j jJelzded ) Ca lego l) ' #5 : > 50 c m #4: > 60 ::; 100c m I 2 3 Qu a lity o f n ew sn o w (v e r y loose" loose , s li g htl y co nso lid a ted , quite co n solid a ted , co n so lid a ted ) slight!)' consolida led #3 4 5 6 Sn ow te mpe r a ture (3 d ) (co ld , co ld --+ w a rm , wa rm --+ co ld , wa rm , w e t ) Pros p ec ti\ "e ch a n ges (s tro n g ly \\"ea ke nin g , sli g htl y wca k e nin g , with o ut a n y inOu en ce , sli g htl y co nso lid a tin g , stro ng ly co n so li- d a tin g ) I n c rease o f sn o w d e pth (3 d ) ( 10, 30 , 5 0 , 8 0 cm ) I:'-ITERPRETATION R e lease pro ba bilit y n e w sn ow: (ve r y lo w , low, m od e r a t e , la r ge , \"e r y la r ge ) EXPLANATIO:\T A t e nt a ti\ "e co nclu sio n fo r this exa mpl e is situ a ti o n :\T o : 170 raid slighl!)' weakening 77 cm la rge? R e lease p ro ba bili ty n ew sn ow: large [t is import a nt to n o te , h owev e r , th a t in th e a bo \ "e c a se n e,," sno " " d e pth is: in s tea d o f: L ess i m po rt a n t differ e n ces a re th a tin th is case: Th e A S SERTION LEVEL fo r t hi s t e nta ti I"C co nclu sio n is: 85 A te nt a ti ve co n c lu sio n fo r thi s exa mpl e is situ a ti o n N o : 4 1 NONE R e lease pro ba bilit y n ew sn ow: l'e!} Lcuge I t is imp o rta nt to n o te, ho ,,"e\"e r, th a t in th e a bo \ "e ca se qu a lit y o f n c \\' snow is: in s teael o r: L ess imp o rt a nt diffe r e n ces a re th a t in thi s ca se: i\O:\,E Th e ASSERTION LE\ "EL fo r thi s te nt a ti\"e co n c lu s io n is: 8 1 A te nt a ti\ "e co nclu sio n fo r this exa mpl e is situ a ti o n No : 86 Rel ease pro ba bilit y n e w sno \\' : hllge It is impo rta nt to n o te, ho ,,"e ve r , th a t in th e a b o \ "e case qu a lit y o f n c \\' sn ow is: in s tea d o f: L ess impo rt a nt diffe re n ces a re th a t in thi s case: Th e ASSERTION LEVEL [o r thi s t c ntati\ "C co nclu s io n is: 8 1 A t e n t a ti\ "C co n c l usio n fo r this exa mpl e is si tLl a ti o n N o : 29 1 R el ease pro ba bilit y n e w sn o w: It is imp o rt a nt to n o t e, h o weve r , th a t in th e a bove c a se n c,," snoll" d e pth is: in s tca d o f: L ess impo rt a nt diffe r e n ces a re th a t in thi s ca se: in crea se o f s n o ll" d epth (3 d ays ) is: in s t e ad o f: Th e A S SERTION LEVEL fo r thi s te nta ti I"C co nclu sio n is: 8 1 NONE modera le # 1 #2 #4": > 50 ::; 80 cm > 50 c m 2: 15 ::; 30 cm sli g h tl )" co nso lid a ted loosc sli g hll y co nso lid a ted quit e co nso lid a ted > 50 C I11 2: 5 :s; 15 c m > 5 0 ::; 8 0 cm 2: 30 < 50 cm p a r a m e te r va lu es (w eat h e r, sn o w a nd sn o w- cove r d a t a ) fo r th e g ive n d ay " Th e log ica l ra n ges a r e, fo r exa mpl e, in th e c a se of th e 3 d s um o f n ew s n o w d e pth 0 """ 10 , 10 """ 3 0 , 30"" " 6 0 , 6 0"" " 120 a nd 111 0 r e th a n 12 0 e l11" Fin a ll y , th e d ec isio n o r inte rpre ta tion is th e d eg ree o f haza rd a nd in m ost \"e rsi ons o f th e m o d e l DA \ 'OS (see below ) th e a ltitud e a nd th e as p ec t o f th e mos t d a nge ro u s slo pes" 32 3 Downloaded from https://www.cambridge.org/core. 06 Apr 2021 at 02:05:53, subject to the Cambridge Core terms of use. https://www.cambridge.org/core J ournal 0/ Glaciolog), 4. INPUT, OUTPUT AND VERIFICATION \\'e h a y e c h ose n our input p a r a meters (call ed qu es ti ons in the jud g m e n t processo r ) fro m a d a ta se t wh ich is beli eved to be represe nt a ti\ 'C o f th e region co n sid e red (Da \·os : a bo ut 50 km 2) : seve n qu a ntities a re m eas ured in th e mo rnin g (a t th e e:-:p erim e nt a l pl ot of SFISAR below \\·eiss Ouhj oc h , 2540 m a.s .l. , o r a t th e littl e p ea k a bo\·e th e in stitute, th e so -ca ll ed [n stitutsgipfel, 2693 m a. s.l. , by th e a ut o m a ti c wea th er sta ti o n o f th e Swiss 1\1e teo rolog ica l Institut e ) , fo ur qu a ntiti es a re pros pec ti\ "C \·alu es for th e d ay co n sid cred a nd ten qu a nti ti es d esc ri be th e ac tu a l sta te o f th e sno\\' cO\·e r b ased on slo p e m eas urements perfo rm ed a bo ut every 10 d. Th ese prin cip a l d a ta a r e give n in T a bl e 2. A d esc ripti o n of th e a yal a n che h aza rd is associated \-,·ith eac h d a ta se t co nsistin g of th e a bo\"C wea th e r, snow a nd snow-cover d a ta. I t see m s m ost ap propri a te to choose as o utput o f a n e:-:p ert sys tem e:-:ac tl y th e stru cture th at is usua ll y used by forecas ters. So th e ass istin g tool ·'s peaks·' th e sa m e la ng uage as th e fo recas ter. The qu es tion thu s is: whi ch d eg ree of haza rd d esc ribes th e prese nt ava la nch e situ a ti o n a nd where is it loca ted ? Onl y th e hi g h es t ex istin g d egr ee o f h aza rd is gi\·e n ; th e loca ti on is give n b y th e slopes, d esc rib ed in term s of a ltitud e a nd aspec ts, th a t a re suppose d to be th e m os t d a ngero us in th e region. So th e gi\·e n d egree of h aza rd is not a t a ll a \ ·e raged o\·e r a ll as pec ts; it is d efinit ely w ro ng to d ea l with aYe rages in thi s co ntex t. Th e refo re, th e a \·a la n c h e h aza rd is fo rmul a ted first o f all as d eg r ee of haza rd (l .. . 7) . Seco ndly , t h e lowe r limit T ab le 2. Princi/Jal daLa lI sed ill the two different D il J ·OS and J10D UL models . D, D ata llsed in the DA J·OS model; AI , Data llsed ill the l\r/OD UL model 1. IVI eaSllrements D , M D , M D , M D , M D , M 1\1 ;,,1 11 . Prognostic data D , ~1 D , M ?If M Ill. Snow -cover data D D M M ~r M M :vr M ~I 324 new sn o w d epth to ta l sn o \\· d epth pen e tr a ti o n d epth wind p eed a nd wind direc ti on ai r te m pe ra tu re sn o w te m per a tu re n ew s n ow d ensit y progn ostic air temp e r a ture a t noo n progn os ti c ind ex of d a il y radi a ti o n progn os ti c mea n wind sp eed pro g n os ti c new sno w d e pth ind ex o f snow-cover s ta bility d epth o f criti ca l laye r res ult o f Rutsc hbl oc k tes t ty p e o f release (RE tes t) ty p e o f critica l laye r (RE test) to ta l sla b thi ckn ess (RE tes t) new sn ow sla b thi c kn ess (RB tes t ) ty p e o f profile (RE tes t ) sn o w d epth a t th e t est site d a t e o f Rutsc hbl oc k tes t of th e prim a ril y end a ngere d a ltitud es is giye n in s teps of u suall y 20 0111. ( > 1200, > 1600 , > 1800. > 2000 . > 2200, >2400, >2500 , >2600, >2800 ma.s.I. ). T hirdl y, th e ma in as pect is d esc rib ed as ei th e r o n e of th e mea n direct ions ( .\", XE . E , SE. S. ,\ ·W ) a nd a n acco rdi n g sec tor ( ±4S', ± 67', ± 90 ) or as extreme slopes o r all slopes . Th e \I·es terl y secto r is no t so fr equ entl y end a n ge r ed a nd if so, o th er as p ec ts a re a lso e nd a nge red , so it m ay be d esc rib ed b y th e o th er o n es. If th e haza rd is giYen, for exa mpl e, as 4, >2400 1110.5.1. , N E± 90°, thi s m ea ns hi gh h aza rd on slo p es with aspec t fr o m n orth wes t to so uth eas t a bo\·e 240 0 m a .s.1. Three exa mpl es a re give n in Fi g ure 6. N s 1, > 2500 m, N ± 45 ° N s 3, > 2000 m, NE ± 90 ° N s 5, > 1500 m, all Fig . 6. T hree e\'amples of hall' the regional avalanche haza rd is described. Downloaded from https://www.cambridge.org/core. 06 Apr 2021 at 02:05:53, subject to the Cambridge Core terms of use. https://www.cambridge.org/core Sc/twei;:.er alld F6itll: A l'ClLa 11 che Jorecast ing - all e.\pert ~)Jstem ajJProac/z V erifica tion Th e "real" ava la nche ha za rd th at Vie use in the dat a base of the DA VOS mod e l to teach the sys te m is th e result of an "a pos te riori " criti ca l assess m e nt o f the h azard , th e so- ca ll ed ve rification (Fbhn a nd Schweizer , 1995 ) . Th e verification h as again th e same stru c ture as th e mod e l ou tput . Oth e rwi se, it is h ard '" poss ible to verify th e a\'a la nch e h aza rd. Se\'e r a l s tudi es 011 the \ 'e rifi ca tion o f th e a \'a la n c h e haza rd with the help of th e so -ca ll ed a \'a la nc h e - act ivi ty ind ex were no t sufTi c ie ntl v su ccessful U ud so n a nd King, 1985; Giraud a nd o th ers, 1987; Remund , 199 3 ). On e reaso n is that in th e case when no a \'a lan c h es are prese nt or o b se rv ed . th e a va la nch e haza rd is no r necessa rily \'e r v low o r even no n- ex is tent. H e nce, it is ob\' io u sly wrong to use th e o b se n 'ed a\'a lan c he ac ti\'it y as th e so le o utput pa ram e te r o f a n ass is ting tool fo r r eg iona l ava lanc he fo recas tin g. Th e ava la nc he \'C rifi ca ti o n IS not id e ntical to th e ,l\'alanche h aza rd desc rib e d in th e public a\'a la nch e warnin g . A s th e warning is prosp ec ti\'e a nd th e d ata ba se m ay b e in co mpl ete a t th e tim e of fo recas tin g. the a\'a lanc h e h aza rd might h a \' e been in fac t la rge r o r sm a ll e r. It is a hl'a ys eas ie r to assess th e a\'a la n c he haza rd in hind sig ht! Op era ti o nall y, th e \'e rifi ca ti o n has bee n d o ne some d ays la te r co n s id e rin g the obse n 'e d a\'a la nc he aC li\'it y ( natura ll y a nd a rtifi c ia ll y released ) , th e pas t weat her co nditi o ns, the a dditi o na l sno w- cover tes ts. th e bac k- co untry skiin g ac ti\'it y a nd se \ 'C ral o th e r , partly perso na l o b se n ·a ti ons. Snow-cover tes ts for m a n impor- ta nt part o f th e \'e rifi ca ti o n lI·o rk. Lik e th e rea l-tim e assess m en t itself, it is an ex p e rt task . W c est im a te th a t th e \'Crification d esc ribes th e ava la n c he situ a ti o n co rrec tl y in a bo ut 90'10 o f th e d ays a nd thu s is much more a cc ura te th a n th e forecas t. L' sin g the d eg ree of h aza rd fi 'o m th e ava lan c h e fore cas t as o utput parameter would o n ly be rea so nab le if the lI'ar nin g w e re a hl'ays co rr ec t. An e rro neo us i nt e rpreta ti o n s hou ld not be e n c losed in th e d ata base. Comparin g th e fo r ecas ted d eg ree of h azard to th e \·e rifi e ation. th e fo recast seems to be co rrec t in a bo ut 70% o f th e d ays . So it see m s ob\· io us that th c use of thi s typ e o f \ 'C rifi ca ti o n rep rese nts a substa ntia l impro\'C m e nt (or the d e vel o pm e nt o f' ex p ert sys tems in the fi e ld of reg io na l a \ 'a la nche for ec a s t i n g . Furtherm o re, \ 'C rifi ca t io n is a prereCJuisite [or c\'aluating th e qual ity of a n y assistin g too l for reg io n a l a \'a lanche for ecas tin g a nd , of co urse, a lso fo r impro \ 'in g th e lI'arning itself: 5. MODELS Us in g the CYBERTEK-COGENSYST\I jud g m e nt pro - cess in g sys te m , we d e\ 'C lo p e d tw o dif'rerent typ es of' mod e l: DA VOS and I\ fOD U L. Th e DA \ ' OS model uses 13 weather, sn ow and sno \\'- co\'e r param e te rs an d e \'aluales th e degre e o f' ha za rd , th e a ltilUd e a nd th e as p ec t. The m ode l is ex elu si\'e ly d a ta-b ased, wherea s th e ~IODU L m od e l is both d a ta - a nd ru le-b ased. It u ses 30 input p a ra m e te rs s tep ll'ise a nd the e \'a lu a ti o n o f the d eg ree of' h aza rd is th e res ult of 11 inter co nn ec te d jud g m e nt pro ble ms that a re fo rmu lated acco rdin g to th e re leva nt processes . Th e sys tem tr ies to m o d el th e d ec ision-m a kin g process o f a n ex pe rt a\'a la n c h e foreca ster. DA VOS IIlodel Gellf1'al Jeatllres : inpllt . OlltPllt and kll owLedge base Th e DAVOS m ode l uses th e input p a ra me te rs g i\ 'e n III Tab le 3. M os t of' th e va lues a r c calc ula ted fr o m nin e prin ci pa l va lu es (Tabl e 2 ) acco rd in g to o ur ex p e ri e n ce . Th e id ea was to ta ke int o accollnt ce rta in rele va nt pro cesses, e.g. n e w snoll' se ttl eme nt. D e ta il s h a \'e bee n g iven in Schw e ize r and oth er s ( unpub li shed ) . I n th e fo ll ow in g so m e o f th e ela bo ra ted p a r a m e ters are bri e fly d esc rib ed. The settLemellt-qllotienl p a ra m e te r co mpares th e in c rease in sn o ll' depth durin g th e las t 3 d wi t h th e s um of' th e nell' snow d e pth o f th e las t 3 d. Th e sma ll e r th e \' a lue th e be tt e r th e se ttl eme nt. H owe \ 'e r , the se tt le m e nt inelud es no t onl y th e new sn ow but a lso the o ld-sn ow se ttl e me nt. Th e co nso lidati o n o f th e surface laye r is d esc ri bed by th e /Jenelraliol1-qlloliml param e te r : th e p e n e t- r a ti on d ep th today di\·id ed b y th e pe ne tra ti o n d e pth yes te rda y . Th e h e at tra nsport int o th e snow cO\'e r is taken int o acco unt by a d eg ree -d ay p a r a m e te r: the slim of th e /Jo silive air tem/Jeralllres alllOOIl oJ Ih e lasl 2 d and the presenl dCI)' ( prospec ti\'ely ) al 200011/ a .s./ . ( the average a ltitud e o f th e reg io n co nsid e r e d ) . Th e snow tr a n s p o rt is includ ed by th e blowing -sl/ ow param ete r: th e s lim of a dditi o n a l wind- tr a nsported snow in leewa rd slo p es OH r th e la s t 3 d ( F b hn and Haec hl e r, 19 78 ) . Th e radialioll illde.\ IS an es tim a ti o n ( I , 2 o r 3 ) o f th e daily lo tal g loba l so lar radiation fo r th e prese nt day. I mean s below th e lo ng - term m ea n \'a lu c (or th e g i\'en d ay, 2 about a nd 3 a b o ve , res p ec ti\ 'C ly. Th e .mOW -CONr slabili£J' illde\ ( I to 5 ) is an es tim a le of th e s tate of th e sn ow CO \'e r co nsid e rin g th e sn ow pro (il es a nd Rutsc hb loc k tes ts th a t a re a\'ail a ble [or th e reg io n. The dep lh oJ Ih e crilical /c£J'er is a n es tim a ti o n fi 'o m th e snow profiles a nd th e Rutsc hbloc k tes ts. W e u s u a ll y di spose of th e sno w profile from th e stud y plot and at least of o ne t y pi ca l sno'" pro (ile with a Rutsc hb lock tes t ri'o l11 a slope in th e D a \'os a rea, th e la tt e r ofte n performed b y o urse h-es. Bes id es th e input param e te rs, w e h a\'e a lso c h ose n th e ran ges for each of' th e input p ara l11 e te rs acco rdin g to o ur T able 3. Inpul jJarameters al/d /ogi{{{/ rallgeJ ./or th e D.·( J 'OS model Inpul /Jaram elers Sum of ne\\' snow d ep th (3 d ) Penetra ti o n d e pth Tot a l snoll' d e p th (3 d befo re ) Sett le m en t quotient Pe netra tion CJ u o ti e n t SUI11 o f bl o win g s n o w (3 d ) Air te mpera ture T e mp e rature diITe re nce Sum o f th e positive te mp era tures at noo n a l2000I11 a .s.1. (3 d ) Ind ex o f radiati o n I nd ex of sno\l' -cove r sta b ility W ind direc ti o n D e pth of cr iti ca l lave r B ollnda ries / choices 10 / 30 /60 fl2 0 CI11 5/15 /30 /50 e l11 70 / 1 00fl50 /200 cm 0.01 /0. 36 /0 . 7/ 1.0 0.4/0.8 / 1.2/ 3 .0 2 / 5 / 1 Ofl5 Clll 15 / 8/-3/0 C 5 /0 /5 / 10 c: 0.0 1/ 3/6 / 10 C I , 2, 3 I, 2, 3,4,5 N W , NE , SE , S W , 00 20 / 40 /60 /90 c m 325 Downloaded from https://www.cambridge.org/core. 06 Apr 2021 at 02:05:53, subject to the Cambridge Core terms of use. https://www.cambridge.org/core J ournal oJ Claciology ex p eri ence . As m e nti o ned above, eac h of th e in put d a ta is assoc iated with o n e of up to fi ve logica l ra nges. Aft er seve ra l yea rs o f d a ta acc umul a ti o n , we ar e fin a ll y ab le to ch ec k w heth e r th e c hose n ra n ges w c rc r easo na bl e o r n ot. T wo exa mpl es b ased on th e 9 yea r d a ta base a r e g ivc n in F ig ure 7. \ Vh e reas th e 3 d sum o f th e new sn ow d epth see ms to catego rize quit e we ll , co m p a red to th e ve rifi ed degree of haza rd , th e set ti e m e n t q uo ti en t sh ows n o sp ec ifi c trend. This is in acco rd 'vV i th th e st ud y o f P erl a ( 1970 ) but it is in co ntras t to th e o pinio n of expe ri en ced fo r ecaste rs. Th a t p roba bl y d oes no t mea n th at th e set tlement qu o ti ent is not imp o rt a nt a t a ll ; it mi g ht be releva nt but o nl y in ce rt ain situ a ti o ns, i. e. co mbin a ti o ns of input d a ta . Sta tisti ca l me thod s, in pa rti cul ar uni var- ia te, d o not tell th e w hole truth. Th e DA VO S m o d e l a lso d oes no t co nsid e r th e set ti em en t q uo tien t as im p o rra n t (T a bl e 4 ) . Th e output res ul t is th e ava la n ch e hazard d esc ri be d as deg r ee of h aza rd , a ltitud e a nd as pect of th e m ost d a n gero us slop es (see a bove ) . Th e kn ow led ge b ase of th e D A VO S mod el co n is ts of th e d a il y d a ta fr o m nin e winte rs ( I D ecemb er- 3D April ), i.e . 136 1 situ a tio n s. Durin g thi s tim e period , 22 si t u a tio ns we re pa irwise id e n ti ca l, i. e. eac h o f th e in p u t p ara m e ters of th e 2 d co nsid er ed belo ng to th e sa me logica l ra n ge. A bo u t 700 000000 situ a ti ons a r e theo reti call y bu t, of co urse, n ot ph ysica ll y p oss ibl e . DijJerent versions: D A I "OS] , D A VOS2, D A VOS3 1/DA VOS32 and DA VOS4 The o ri gin a l versio n of th e DA VO S model w as call ed DA VO S I. Th e ex p eri ence wi t h thi s ve rsion h as g ive n r ise to fu rt her ver sio n s. T he \'alu es of th e logica l imp o rt a n ce of th e o ri g in al D AVO S I ve rsio n (T a bl e 4 ) show cl ea rl y th at this ve rsion is hardl y a bl e to disc rimin a te well. T welve of th e 13 input pa ra m eters a re co n sid e red as m ajo r o n es. As a con seq uen ee, loo ki ng for si mil ar si tu a ti ons m ea ns th at te n of the prese n t-d ay inpu t par a m e ter va lu es h ave to be in th e same ra n ge as the past-d ay va lu es. T hi s r eprese nts a ve r y st ri ct co nditio n for th e simil a rit y a nd r es ul ts in a 7 r---~----------~---------------, 50 100 150 200 250 3 d sum of new snow de pth [c m] 7 ~------------~--------------------, "0 6 - - • - ~ .. + - - .. :- .. - - • - - - - ~ co: 5 -------. -:- 0 .•• -;... -;- --- - ~ '0 4 :.- . .. ... --. . ......... ~- ~ 3 ____ "'!II!!'O ••• - - -. - Cl> .....--: {l 2 . "~=::::':....;.~ V 1 ~ __ ~~ __ ~~~~--+-----~-----4 o 0.5 1.5 s ettleme nt coefficient 2 2.5 Fig . 7. Comparison oJ the 3 d slim !if the new SIlO W depth (a bove) and the settlement quotient (below) with the deg ree of hazard to check whether the logicaL ranges chosen categorize the data appropriately . Average degree oJ hazard Jar each category is also shown . T he da ta base consists of 136/ situatiolls from nine winters. la rge numb er of unin terp reta bl e situ a ti o ns. Thi s fac t see m s d e finit ely to h e du e to th e d esired o utpu t res ult th at co n sis ts of three ind ep e nd ent a nd equi va lent co mp o n ents (d eg r ee of haza rd , a lti t ud e a nd as p ect ) . In th e case of different ind epend ent output res ults, th e CYBERT EK- CO GEN SY SH 1 judgment processo r offers the poss ibili ty or choos in g one of th em as th e d omin a nt ou tp ut r es u I t. H ence, we use d a seco nd ve rsion or the mod el D A V O S: DAV O S2. 'vVh ereas in the D A VO S I \'e rsion a ll three res ul ts are eq ua ll y importa nt, in th e D A VO S2 ve rsion T able 4. Values oJ lite logical imjJortance oJ the different versions of the D A VOS model. B old figllres indicate so -called maj or parameters InjJut parameters D AVOS l D AVOS2 DAVOS31 DAVOS32 D AVOS4 EHN < IDem ) EHN 2: ID el11 S um of new sn ow d ep t h (3 d ) 0 100 0 100 100 Penet rat ion d ept h 83 28 14 47 29 T o ta l snow d epth (3 d before) 83 65 79 60 65 Se ttl ement qu o ti ent 50 2 1 15 18 19 P e n e t ra ti on q uo ti en t 41 24 18 36 27 S um of b lowin g sn ow (3 d ) 66 33 35 41 32 A ir te m pe ra ture 66 23 22 18 15 T e mp era ture difference 24 15 7 17 13 S um o f the positi ve temperatu res a t n oo n a t 2000 m a.s. 1. (3 d ) 41 29 30 4 22 In dex or ra di a ti on 44 II 17 II 16 Ind ex of snow -cove r s tab ility 100 86 81 70 77 Wind di rec ti o n 33 26 26 18 2 1 D epth of criti ca l laye r 79 51 100 47 83 326 Downloaded from https://www.cambridge.org/core. 06 Apr 2021 at 02:05:53, subject to the Cambridge Core terms of use. https://www.cambridge.org/core Sclill'ei::er and Fiilzn : .. h'a/allclie forecasting all e.\/Jnt sjlstem ajJProacli th e first o ut p ut res ult , th e d eg ree of haza rd, is th e d o min a nt o ne. Thi s lea ds to different \'a lues of th e logica l impo rt a n ce a nd acco rdin gly to d ifferent int erpre ta ti o ns. ~los t re le\'a nt is th e sm a ll numb er (fo ur ) of maj or p a ra m e ters a nd he nce th e be tt er se lection pe rfo rm a nce: ha rdl y a n y unint erpre ta bl e situ a tio ns, In th e D A VO S2 \'e rsion , th e va lu es of th e log ica l impo rta nce see m to be close r to ge nera l ex peri ence th a n to th e DA VO S I \'e rsion ", he re, for exa mpl e, th e ne\\' snow d e pth has no impo r tance a t a ll, Th e reaso n is t he so n of ou tput resu lt a nd th e predo min a nce of situ at ions wi th no new snow , Fro m ex peri e n ce, it is o b\'io us th a t it is quite imp o rt a nt w h e th e r fo r a g i\'e n day th ere is n e w sn o \\' o r no t. H e n ce, wc tri ed to ta ke int o acco unt thi s fac t b y d e\'e lo pin g t\I 'O n ew ve rsio ns: o n c fo r th e mo re fi-cqu e nt situ ations with o ut sn o \\'fa ll a nd th e o th er fo r th e situ at io ns \I' ith sno \l'fa ll - th e DA \'O S3 l a nd D A \ ' O S32 \ 'e r sio ns, r es pec ti\'e ly, Th ese two \'e rsio ns co ncc ntra te o n th e d eg ree o f h aza rd lik e th e DA VO S2 \'e rsio n. Th e kn o wl ed ge base o f th e D A VO S3 1 I'C rsio n co nsists o f a ll d ays fr o m t he las t nin e w int e rs, w hen th e 3 d sum of ne\\' sn o w d e pth was less th a n 10 c m ; th e co mpl e m e nt a r y se t o f d ays w ith th e 3 d sum o f ne \\' sn o\l' d epth cC] u a l to or la rge r th a n 10 cm fo rm s th e kn o wled ge base o f th e DA VO S32 \ e rsio n. Th e d iffe re n ces in th e log ica l im po rt a nce in th e tw o \'e rsio ns a rc quite t y pi ca l (T a ble 4 ). Th e \'a lu es o f th e log ica l imp or ta n ce of th e DA \ 'O S3 1 (n o n e \l' sno\l' ) \'ers io n a re simil a r to th e log ica l imp o rt a n ce o f th e o ri g in a l DA VO S I \'e rs io n , w h e r eas th e log iea l impo rt a nce of th e DA \'O S32 ve r sio n (n ew sn o\l' ) a re simil ar to th e o nes o f th e D A \ 'O S2 \'e rsio n . Th e num be rs o f m aj o r pa ra m e te rs a rc fo ur a nd seve n , res p ee ti\ 'C ly, fo r th e D A VOS 3 1 a nd DA VOS 32 \'c rsio n s; so th ey sho uld di sc rimin a te q uit c we ll, Fin a ll y, th e o utput r es ult \l'as redu ced to the d eg r ee o f h aza rd : D A VO S4 . Th e D A \ ·OS 4· \'e rsio n is m os t a ppro pri a te fo r co mp a ri so n \I' ith s imil a r fo recas tin g m o d e ls a nd \I'C h o p ed th a t du e to th f' sin g le ty p e of' o utput th e DA \ ' O S 4- \'(' rsio n sh o uld di sc rimin a te b e tt e r th a n th e o th e r \'e rsio n s, In a ll th e diffe re nt I'C rsio ns th e input pa ra m ete rs d esc ribin g th e Sla te of th e sn ow cOl'e r pro \'ed to b e imp o rt a nt (T a ble 4 ) , MODU L n lOdel General featu res all d structure Th e ex p e ri ences w ith th e di ffe re nt \'e rs io ns of th e DA \ ' O S mod e l, d esc rib ed a b Ol'e, di rected u s to tr y a differe nt , m o r e d e te rministi c a pproac h, Ori g in a ll y, ~IT hoped th a t th e DA \ ' O S m o d e l. based o n th e jud g m e nt p rocesso r , \l' o uld be ab le to c h oose th e re le \'a nt p a ra me te rs fro m th e 13 input pa ra m e te rs a cco rdin g to th e situ a ti o n a nd ge n e rall y so m e h o \\' to recog ni ze th e hidd en stru c ture o f reaso nin g behind it. D esp it e th e sa ti sfac to l'\' res ults o f so m e of th e \'e rsio n s o f th e DA VO S m od el (sce sec ti o n b e low ) , it sec m s that thi s aim w as too a mbiti o u s; th e pro blem seems to b e too co mpl ex or th e m e th od no t good e n o u g h, So \I'e d ec id ed to " h e lp " th e sys te m b y s tru c turin g th e input d a ta. Th e d es ig n of th e :-IOD U L m o d e l is th e re fo re q uitc simil a r to th e \I'ay a prag m a ti c ex p e rt fo recas te r d ec id es ( Fi g, 8 ) . Firs t of a ll , it is d ec isi\'e ,,' hc th e r th e re is ne\\' sn ow o r no t. Eith er th e ex p e rt h as to assess th e n ew -sno \\' s tabilit y o r h e /sh e di rec tl y assesses , \\' ith o ut n e w sno\\', th e o ld- w eath er. snow and snow cover data Fig , 8, S tructllre qftli e MODeL model : 11 sub -problems alld their relatioll , Shaded bOles are 011£), considered in tlie ((oe (if lIel(, SIlOl('. sno \\' s ta bili ty w hi c h is o ft e n simil a r to th e stabilit y I cl befo re, exce pt if the re is, fo r exa mpl e, a large in c rease o f hea l tr a n s p o rt a nd /o r ra di a ti on , So h e / s h e s tru c tures t h e inpu l d a t a acco rdin g to th e dine- re nt ste p s in th e d ec isio n p rocess, I f bo th th e n e \\'-s n o w sta bilit y a nd th e old-sno \\' sta bilit y, in c ludin g b o th th e efIec t o f th e ,,'ea th e r as fo recas t fo r tod ay, h a \'c b ee n assesse d. th e t\l'O re lease p ro b a biliti es a rc co mbin ed. T a kin g int o acco unt th e effec t or th e te rra in a nd o f th e s kier as a tri gge r , th e d eg ree o f haza rd is fin a ll y d e te rmin ed, At th e m o m e nt, o nl y th e d eg ree o f ha za rd is g i\'e n ; th e a ltitud e a nd th e as pec t o f th e l1l os t d a nge ro us slo p es, as g i\'e n in m os t o f' th c \'e rsio n s o r LI lt' DA \ ' O S mod el, h ",'e no t )Tt b ee n impl eme nt ed. Sub -j1m blellls Eac h o r th e sub-pro bl e m s as , for exa mpl e, qllali{J' qf lIew SilO1£' o r stabili!J' DJ old SII01I' re prese nt s a jud g m e nt probl e m , as d esc rib ed a bo\'e, a nd is hence prin c ip a ll y stru c lured lik e t h e D A \ 'O S mo d e l. Th e d iffe re nt s ub-pro bl ems a r e just s m a ll e r th a n th e D A VO S model , i. e , co nsist o f o nl y three to e ig ht input p a r a m e te rs, O fte n. o nl y three o f th e input p a ra m e ters a rc co n sid e red as m aj o r pa ra m ete rs. Thi s is a g rea t ach 'a nt age, since a mu c h sm a ll e r kn o \\'led ge ba se is suffi c ie nt to ob ta in good int erprc ta - tio ns a nd th e SI's tem us u a ll\' lea rn s fas te r a nd bette r fr o m , , th e log ic b e hind th e d ec isio n process , S ub-pro blem s w ith o nl y a b o ut fi\ 'e input p a r a m e tcrs mos t o f t h e tim e m ay find a simil a r pas t situ a ti o n that is id e nti ca l to th e ne w situ a ti o n , b ased o n a kn oll' led ge base o f o nl y a bo ut 10 0 situ a ti o n s. i mjJ/icit I'll les It is e \'e n p oss ible no t o nl y to build up th e kn o \l'led ge base \I'ith r ea l situ at io n s but a lso to co n Slru c t rea li s ti c situ a ti o n s b y \'a ryin g th e m aj o r input p a ra m e ters in a reaso n a bl e se nse. Thi s is im poss ibl e in th e DA \ ' O S mod e l. S o, if th e ex pe rt fee ls sure a bo ut o n e o f th e sub- pro bl e m s o n th e innu e n ce o f o ne o f' th e input pa ra m e te rs, may b e in co mbin a ti o n \I' ith a no th e r o n e, he/shc m ay sys te m a ti ca ll y co nstru c t rea li sti c situ a ti ons a nd d ec id e sys te m a ti ca ll y , But thi s m ea ns no thi ng o th e r th a n in c ludin g a rul e, n o t ex pli c itl y but impli c itl y , An exa mple o r such a n impli c it d cc isio n rul e used in th e 327 Downloaded from https://www.cambridge.org/core. 06 Apr 2021 at 02:05:53, subject to the Cambridge Core terms of use. https://www.cambridge.org/core Journal oJ Glaciology T able 5. Gel/eral Tule to decide on the degree of lza::.ard in the sub-jJroblell1 fin a l m e rg in g; principally dep endent 011 the co mbin ed (na tura l) r e lease pro b a bilit y and the infl u e n ee of th e ski e r , but also de/mulml all lhe overa ll c ri ti ca l d ep th by lhe po telltial avalanche si::.e and volume and 0 1/ lhe d e pth of s ta bl e o ld sno w b)i the lerraill roughness Influence oJ skier L o w ?-.l od er ate Hi g h Combined release probabilil)' r 'e ly Low lvIoderate High low I 2 I 2 3 2 3 4 4 4 4 V a lid if o " e r a ll c ritica l d e pth H cril = 15 . .. 5 0 c m , else, if H"'il < 15cm.l d eg ree o f h azard lessor ifHcril ;:::50c m a nd com bin ed rel ease p ro b a bilit y hi g h , th e n d egree o f h aza rd = 6 or 7 a nd if d ep th o[ sta bl e o ld sn o w H bouncl > 6 0 c m , e lse, if H bound = 30 .. . 60 cm , th e n I d egree o f h aza rd less o r if H boLl ll cl < 30 cm , th e n 2 d eg rees o f h aza rd less exce pt i[ co mbin e d rel ease p ro b a bilit y = moderate, n o redu c ti o n of h aza rd , o r i[ co mbin ed re lease prob a bilit y co nsid e ra bl e, o nl y I d eg r ee of haza rd less . su b-pro bl e m Jinal merging is g i" e n in T ab le 5 . Thi s is, o [ co urse, r a th e r ex ha ustin g w o rk bUL th e a d va ntage is th a t on e is m o re fl exibl e in o n e's d ec isio n th a n in th e case w he re o n e uses a stri c t ex pli c it rul e. It is easy fo r exa mpl e to includ e no n-lin ea r rela ti o n s. Furth erm o r e, it is poss ibl e to co n s tru c t ex trem e but s till rea li sti c situ a ti o ns th a t a re usua ll y r a re but o f co urse v e r y imp o rta n t. S o o ne o f th e di sad va n tages o f prin ci p a ll y sta tisti c (o r d a ta )-b ased m o d els us in g rea l d a ta m ay be ovc rco m e. Fin a ll y, o n e a rri ves a t a knO\d ed ge b ase th a t is a mixture of rea l, hi storic si t u a ti o ns d ec id e d acco rdin g to th e " erifi ed h aza rd a t th ose tim es a nd rea li sti c si t u a ti o n s direc tl y d ec id ed acco rdin g to ge n e r a l kno wled ge a nd ex peri ence . Th e pro bl e m is to h ave th e a ppropri a te mi x ture. InpuL parameters Thirty input pa ra me te rs (T a bl e 6 ) a re u sed in II sub- probl e m s inte rconn ec ted p a rtl y b y rul es . Alread y, to o bta in so m e of th e d a ta, a use r Wilh ce rt a in skills a nd ex peri e n ce is required. Th e o utput res ult o f a s ub-pro bl e m is usu a ll y used as a n input p a ra me ter to a noth e r sub- prob le m th a t a pp ea rs la te r o n in th e d ec isio n-m a kin g process . ?-.1 a n y o f th e input-p a ra m eter va lu es a re ca lcul a ted usin g rul es th a t d epend th e m se lves o n th e input valu es . Th e overall critical depth [o r exa mple d epe nd s, a mon g o th e r things, o n th e 3d sum of blowing-sll ow dep Lh th a t is o nl y con sid e re d in ce rta in situ atio n s wh e n sn owdrift is lik ely. NI odifica tions Du e to th e m odular stru c ture, it is eas il y possibl e to m odify a n y o f th e s ub-probl e m s. Additi o n a ll y, th e rela ti ve ly sm a ll numb e r o f input p a r a m e te rs in eac h sub-pro bl e m e na bl es th e kn ow led ge base to a dapt qui ckl y to a n y m o difi ca tion , as [o r exa mpl e a ddin g a n ew input 32 8 T able 6. inp ut jJarameters used ill the Ai 0 D U L model . T he data are grouped according to the availability, i.e . how UIS} it is to gf't thp rlata A . Conventional data New snow d e pth S um o f n e w snow d ep th ( 3 d ) D ensi ty o f n e w snow Snow d e pth C ha nge o f sn ow d epth (3 d ) Coe fIi cie nt of se ttl eme nt (3 d ) Pene tra tion d e pth Coe fIi cie n t of penetra ti o n d e pth Snow te mp e r a ture M ean wind sp eed (3 d ) Sum of blowin g snow (3 d ) Air te mp e r a ture T emp e r a ture differen ce B. Prognostic data N ew snow d e pth in th e eve nin g T e mp e rature d eve lopment until noon M ea n wind sp eed until tom o rrow Progn os ti c ind ex o f ra dia ti o n fo r tod ay C. S/Jecial snow-cover daLa R es ult o f Rutsc hbl oc k tes t T ype o f r e lease (Rut schbloc k tes t ) T ype o f c riti ca l laye r (Rutsc hbl oc k tes t ) T o ta l sla b thi c kn ess (Rutsc hbl oc k tes t ) Ne w sn o w sla b thickn ess (Rutsc hbloc k tes t ) T ype o f ram pro fil e (Rutsc hbl oc k tes t) Age o f Rutsc hbl oc k tes t C ha n ge of sn ow d epth sin ce Rutschbl oc k tes t C riti ca l d e pth of ne\\' snow sla b C ri li cal d e pth o f old snow sla b O ve ra ll c ri ti ca l d e pth EfTcc ti ve c riti cal d epth for s ki e r tri gge rin g De pth o f s t a bl e o ld sn ow p a ra me te r. So th e impo rta nt s ub-p ro bl em influence oJ the skier is stea dil y improved a cco rdin g to th e r es ults o[ sp ec ifi c stud y on sla b- ava la n c h e release trigge r ed b y a ski er (Sc h weize r , 1993 ) . [n th e s u b-pro ble m sl1ow-jJ70Jile analysis, th e snow pro fil e with Rutsc hbl oc k tes t is ro ug hl y interpre ted , a n a im th a t would ac tu a ll y need a n ex pert sys tem itse lf. Ei g ht prin cipal v a lu es (T a ble 2 ) a re used ex clu sive ly for solving thi s sub-pro bl em. T oge th e r with th e sub-probl e m sLabili!), of old snow, it should s ubstitute th e mos t import a nt input p a r a m e ter index if snow-cover stability in th e DA VOS mod els. S o thi s sub-probl e m is a lso und er pe rman e nt improve m e nt. R ece ntl y, t)ipe of release and th e quali{y oJ the critical la)ier o f th e Rutse hbl oc k tes t we re in trodu c ed. Operalionaluse Th e mod e l has to be run int e ra ctive ly by a n ex p eri e nced use r. Th e mod e l sto ps if th e pro p osed d ecision in o ne o f th e sub-pro bl e m s d oes no t h ave a hi g h level o f co nfid e nce, a nd th e use r h as to co nfirm th e d ec isio n befor e th e m od el co ntinu es to run. In th e ex ampl e g ive n in T a bl e I , th e interpre ta tion of th e release /) To babilil)! oJ new snow b y th e Downloaded from https://www.cambridge.org/core. 06 Apr 2021 at 02:05:53, subject to the Cambridge Core terms of use. https://www.cambridge.org/core Schweizer and F61111: Avalanche Jorecastillg~ all expert system ap/Jroac/z Table 7. Qj/OIiI)I requirements Jar determinillg th e pelJonna nce cif the D AVOS mo del Qualit)l D eviation : degree D evia tioll: alt iLude Deviation : aspect of hazard Good 0 ±400 m About ri g ht F a ir 0 ±400m Not co mpletely wro ng 0 \ '\Trong (a ny res ult ) Wron g (any r es ul t) ±l ± 400 m Abo ut ri g h t P oor ±l > ±400 m N ot compl ete ly wrong ± I > ±400 m Wro ng (any r es ult ) Wron g > ±l (any res ult ) (any r es ult ) sys te m IS large?, w hi c h m ea n s th a t th e sys te m is not co nfid ent a nd the int era ctive run wou ld stop. A s a n exp la nation, four simil ar situations are give n w ith th e in te rpr eta ti o n s large, lIel} large, large and moderate . Comp a rin g th e pr ese nt case w ith th e fir st simil ar s itu a ti on indi ca tes ra th er vel)' large th a n l{JIge as o utput for the prese nt situ at ion; th e seco nd a nd third simil ar situations indica te th e output is b e twee n large and vel)' large; th e fourth situ at ion is too far from th e prese nt case to be co nsid e red. So, based on simil ar situation s, the use r wo uld pres um ab ly c h a nge th e int e rpretation to ve l)! Large. Hmve\'e r, th e int erpre tation proposed by th e syste m , large , wou ld no t be wrong bu t vel~Y large see ms to b e more consistent w ith the pre se nt kn owledge base , Th e final output r es ult , th e d eg ree of haza rd , is w ell ex pl ain ed by th e output res ult s o f the diOere nt s ub- proble ms. If th e mod el proposes a differe nt d eg r ee of h aza rd than th e u se r h as ind e pendentl y es timated , the diffe re nc e us uall y b eco m es obvious by in spec tin g th e output res ults of th e sub-problem s. Du e to thi s featur e, th e use r exp e ri e nce s th e mod el not as a bla ck-b ox syste m (d es pite th e prin c ip a ll y unknown a lgorithm ) but as a r ea l supportin g too l to th e forecaster . Th e interac tive u se of th e mod el prov ed to b e \'e r y in st ructiv e. 6. RESULTS Thank s to ve rifi ca tion data, it is poss ible to r ate our m od els quit e obj ec tiv ely, co mparing th e model output da y-b y-d ay to the verifi ca ti o n. Durin g th e last 5 yea rs of opf'ra ti ona l use, th e kn ow ledge base has con ti n u ous ly in c reased. Sin ce th e pe rforman ce of th e mod els d e pend s s tro ngly on th e s ta te of th e kn owledge base, th e r es ults a re not homoge neo us. Thi s is es p ec ia ll y tru e for th e firs t wint e rs with vers io n s of th e DAVOS m od e l. For co n sistency b e tw ee n th e diITe ren t mod els and \'e rs io n s, w e w ill on ly prese nt in th e following th e performance res ults of three winters ( 199 1- 92, 1992 - 93 a nd 1993- 94 ) . DA VOS nlOde } To co mp a re the int e rpre tat io ns provided by th e sys te m with th e ve rifi ca tion , the require ments of quality (four classes: good, Jair, /J oor, wro ng ) (Tabl e 7) were defined . If th e verifi ed aspect is e.g . AE ± 45'" , the rating in th e following cases]v' ± 67>, N Il ' ± 90° and S ± 90° wou ld b e abo1lt rigId, not compleleL)1 wrong and wrong, res pecti\·ely. Cons id er in g the d eg r'ee of' hazard , the a ltitud e a nd th e aspect th e DA VOS I and DA VOS2 ve r sions hav e on average a performan cc of about 65 and 70 % good orJair (see T a bl e 7 for d efiniti ons ) int erpretations , res pec tivel y (Tabl e 8 ) . To be ab le to co mpare the res ults of the DAVOS 1 and DAVOS 2 versions to the res ults ofdilIer e n t sys te m s, it is more co n ve ni e nt to co nsid er on ly th e d egr ee of haz a rd. In that casc, in 52 and 54% of a ll situ atio n s, res p ec tively , for DA VOS I a nd DA VOS2 , the deg ree of haza rd "v as co rrec t co mp a r ed to th e ve rific a tion. 86 a nd 89% of a ll situation s, res p cc ti\'e h' , a re co rrec t or d eviate ± I d eg r ee of hazard fr o m th e ve rili ca tion. Table 8 PelJormante oJ th e D/l I 'OS 7 and DA rOS2 ve rsiolls considering all Ilzr ee Ollt/Jllt results: the degree oJ ha z ard, the altitude alld tile as/leet. ,\ Jean valll es ( proportions ) oJ tlte la st Ihr ee zt'in ters ( /99/- 92 to /993- 94 ) Jor tlte different qualities defin ed ill Table 5 are given QjJaliljl oJ res1Ilt DJ flGSJ DA I'OS2 Good 0.3 5 0.37 Fair 0.32 0.33 Poor 0.19 0.19 Wron g 0.07 0.04 No int erp retation ( n . i. ) 0.07 0.07 As th e o utput res ult no illter/Hetation is cons id e red as neith e r ri g ht nor wrong, th e performan cc ma y be g iven considerin g on ly th c inte rpreted situ a tion s. In that case , the proporti on of co rre c tl y int erpret ed situ ations (degre e of h azard ) is 55 a nd 58 %, r esp ec tivel y, for DAVOSl a nd DAVOS 2. Th e qualit y of these vers io n s ma y dilfer from one winte r to ano th e r b y abo ut 5%. An exa mpl e of th e performan ce durin g a whole winter is give n in Figure 9. The average p e rc e n t age of cor r ect in te rpreta tion s (co n sid er in g onl y the degre e o f' d a nger ) durin g the wint e r of 1993 - 94 for th e DAVOS 2 vers ion was 56% . 329 Downloaded from https://www.cambridge.org/core. 06 Apr 2021 at 02:05:53, subject to the Cambridge Core terms of use. https://www.cambridge.org/core Joumal oJ Glaciology 7 "E 6 u c Q) ::1 C' ~ -Q) > ''-:; I1l ~ 0 .8 1 0 -30 -2 -1 . 0 G'l1 02 0 3 0.6 0.4 0 ,2 0 J[-J~ -- ~ J h j R In. ~ ~ I NEX. M OD NEX. MOD DAVOS 1 D AVOS2 DAVOS4 MODUL BULLETIN rounded Fig , 11 , CO llljJ arisoll of the jJeljo1'lI7allce of the statistical forecas t model J'v,/!'X_ilIO D , of 01l1'fOlll diffe ren t forecast models DIII 'OSI. DAVOS2, DA l 'OS4 al/d MOD UL, and q/lhe /Jll blic lI '({millg BC LLET!. \' dl/ril/g the thlee l(lillters 199 1- 9210 1993-9-1, Th e relative freqlle /l C)1 of tlt e del'ialioll from tlte verified degree of ha::.ard ill th e DOl'OS area 1.\ gIVen, 7. CONCLUSIONS 1' ll e C\1BE I~ rl'EK-COGENSYST~ 1 ' I I '\. JUc g m e n t pro cesso r, foll ow in g th e id ea of indu c til'e d ec ision-mak ing, prOl'ed to be use ful so ftw a r e fo r d el 'e lo pin g spec ifi c a ppli ca ti o ns in th e fi eld 0 (' ava la n c h e-h aza rd assess m e nt. Us in g weather, sno\l' a nd snow-cOl'er d a ta as input parameters, th e d e,'eloped mod els el 'a lu a te th e a\'a lan c h e haza rd fo r a g i,'e n reg ion, Th e n e\l' f'ca tur es a re th e c h o ice o f ela bora te input parameters, es p ec ia ll y more sn ow -cOl'e r d ata, th e non- lin ea r ca tego ri za ti o n o f th e input data, lh e s p ec ifi c a lgo rithm for th e sea rc h fo r simil a r situ a ti ons a nd fin a ll y th e co ncise o utput r es ult. Th e ava la n c h e h aza rd is d escrib ed as d eg rce of h aza rd , a ltitude a nd as pect o f th e m os t e nd a nge red slopes, for th e first tim e acco rdin g to th e sca le used in the fo recas ts, Thi s so rt of o utput res ult is m os t e ffi c ie nt for th c purpose of a l'a la n c h e fo rccast ing; it is mu c h m o re ap prop ri a t c to th e prob lem th a n, for exa mpl e, th c o utput "al'a la n c h c / non- antl a nch e d ay" , Th e use o f o b- se n 'at io na l a l'a lanc h e d a ta a lo n e is in suffi cie n t fo r bot h fo r ecas tin g an d ve rifi ca ti o n, Th e g i I'en o u t pu t resu I t is p ossibl e due to th e effo rt of pe rm a n en tl y I'erifyin g t h e ava la nche haza rd, V erifi ca ti on is thc m os t st rikin g feat ure a nd co mpl e tes th e d a ta se t - al th e prese nt time nin e w inte rs o f wea th e r , snow a nd sn o w- cOl'er d a ta with a co rrespo ndin g l 'C' rifi ed de g ree o f h aza rd - pro babl y a u niq ue se ri es, Th e sn O\l'-CO I'e r d a t a p ro l'ed to be I'e r y import a n t. Actually, it is well knOlI'll th a t a l 'a lanc h e fo recas tin g d e p ends strong ly o n th e sta te o f th e s n o \l' cOl'e r. H o w e l 'e r , a p a rt fr o m th e Fre n c h i\IEPRA m o d e l, until now h a rdl y a n y of th e present m o d e ls halT t a ken in to acco unt this o b v io us fac l. fl l cC lun g a nd T\\' ee d y ( 1994 ) introdu ced th e sn o\l' (,01,(, 1' so m e h ow impli c itl y in th eir mod e l b y co mbinin g the es timat es o f' th e m o del a nd o f the expert. Of co urse, thi s so rt o f dat a is n o t eas il y ava il a bl e but it is a n illu sio n to expe ct th a t a supportin g too l without a n )' sn o\\' -col'(' r d a t a is as p owe rful as th e expe rt for ecas t e r, Th e present-day m e teo ro logy pl ays a n im portant ro le but m os t o f the tim e it is n o t th e d ec is il'C o n c, Th e int eraCl il'e u se of th e m o d e ls pro\'ed to b e a s ubs ta nti a l advantage a nd es pec ia ll y th e \IOD U L m o d e l is I'(' r y in stru Cl il'C , It is I'er)' a pprop ri a te fo r the tra inin g o f juni o r fo recas le rs \I'ith a ce rtain b as ic kn o\l'led ge , Th e m o d c l run b y a juni o r fo recas te r m ay ac hi el'e a bout th e sa me performan ce (a bout 70% ) o n averagc as a se ni o r ex p e rt fo rccaster u s in g th e cO lll 'e n ti o n a l m eth od s . Th e DA \ 'OS m od el, a dat a -b ased mode l. and the flIOD U L model, a co mbin ed d a ta - a nd rul e-b ased m o d e l, h al,(, ac hi cI'('d a pcrformancc of a b o ut 60% a nd 70- 75%, r es pec ti l'c ly, Th e r e ex ist no co mp a rab le o r similar r es ults, b ased on a lo n g -t e rm operat io n a l test, o f a n y diffe r e nt sys te m for th e forecasting of th e reg io n a l al'a la nc he h aza rd, H OII'CI 'cr, th e p e rform a nce of a sys te m ca nn o t be full y d esc ribed b y a n al'erage pcrce nt age o f co rrec t int e rprct- a ti o ns co mp a red to ve rifi ca ti on ; th e pe rfo rm a nce in c riti ca l situ a ti ons is d ec isil'e , In such situ a ti o ns, a ll prese nt m o d e ls a re still no t good e n o ug h, In el'a lu a tin g the pcrfo rm a n ce of o ur sys tems, lI'e hav e bee n quite ri go ro us. Of co urse, a d ev ia li o n of o n e deg r ee o f haza rd d oes no t m ea n th e sa m e in a ll situ a ti o ns; it d e pe nd s o n th e d eg r ee of' haza rd a nd o n the Li irec ti o n of th e d el 'ia ti on, H owe l Tr, in co ntra st t o th e pre fe rence of so m e a\'a la ncht: -fo rc cas lin g se n 'ices , we think th a t avalanche warning is onl y e ffi c ient and fa ir in th e lo n g term ifn o m a rg in o f se c ur it )' is in c lud ed in th e fo recas t. W e a dmit th a t, in a sp ecifi c criti ca l situ at io n , a warn in g that is, for exa mple , o nc d egr ee abol'c th e la tt e r o n I'e rifi ed h azard m ay help prevent accidents , But , if th e fo recas ted d eg r ee is usu a ll y LOO hi g h , th e lI'a rnin gs become in effi cie nt a nd th e warning sr n 'ice I,,-ill soo n lose its c redibilit y , S o, in co n c lu sio n, liT co nsid e r a c!el'ia ti o n o f o nc d egree e ith e r 33 J Downloaded from https://www.cambridge.org/core. 06 Apr 2021 at 02:05:53, subject to the Cambridge Core terms of use. https://www.cambridge.org/core J ournal oJ Glaciology up or down as a \\To ng d ec ision , ind ependentl y of th e d eg ree of haza rd . Th e nex t step in th e d e\'elopm e nt will be to appl y th e mod els in dilTerent region s to assess th eir p erforman ce. Addition a ll y, several of th e sub-p roblems will b e furth er improved and it is also planned to det erm ine th e a ltitud e and th e as p ec t of th e most d a n ge rous slo p es in the MODUL model. Th e corres ponding sub-probl e m s still h ave to be developed. Fin a ll y, th e hazard of wet-snow ava lanc h es in spring tim e will be tak en into acco unt more sp ec ifica lly . The MODUL mod e l co ntain s g rea t potenti a l [or future developm ents. ACKNOWLEDGEMENTS W e thank th e form er direc tor of o ur in stitut e, C. Ja cca rd , for hi s enco uragement at th e start of this stud y. U . Guggis- berg, hea d of I NFEXPERT th a t se ll s th e so ftwar e used , promot ed th e work with co nsid e ra bl e enthu siasm. Variou s memb ers a nd tra in ees from SFISA R assisted durin g fi eld wor k, data management and o peration al use. \ Ve a lso th ank C. Pili s, who co ll aborated in th e ea rli es t ph ase of th e proj ec t. \\'e are grateful to the two anonymo us re\·iew e rs and , in pa rticular , to th e scientifi c editor D. M.l\,I cC lun g; th eir criti ca l co mments helped to improve the p a per. REFERENCES Ar mstro ng, R. L. . E. R. LaC hapcl lc. i\f.J . BOI·is and J. D. h -cs. 1974. Daeio/)!lIflll oJ lI1elhodologr Jor emlualioll alld /Hediclioll of m'alallche IlG~ard ill Ih e SIIII JIIIIII 1II0ulllaill I/rell of soulhu.'"lem Colorado . Bo uld e r. CO, Un i\'C rsit )' of Co lo rado. Institute of Arctic and Alpine R esearch. INST. \ ,\R 14 -06 -7155 -3. ) Bo is, P .. C. Oblcd a nd W . Good . 1975. \! ulti\'a riat e data analysis a s a lOo l for d av -b y-d av avala nche fo recas l. IlIlemllliolllll Associal ioll 0/ H)'flrological Sciellces Publicalioll 114 (Symposium a t Grindcl\\'al d 197+ Sno\\' i\l ec hani cs ), 39 1 403 . Bo lognes i, R. 1993 a. Artificial intrlli gc ll cc a nd loc al anl lan e hc fo r ec a stin g: the sys te m A \ ·A LOG. III Proceedillgs . I lIl efl/llliolllll ElIIl'I'gellSJ' IIl/d E ngineerillg ConJerellCl' , . Irlillgloll. 1':1. Soc. of Computer Scie nce (S.C. S. ) , San Diego, CA , 1 13- 11 6. Bolog nes i. R. 1993b . :"IX-LOG: idees di recrrices . . \ 'eige; el :it'lI lallches 63, 22 2+. Brun. E .. E. \f art in , \' . Sim on, C. G endre and C. Co leo u. 1989. An e ne rgv a nd m ass model of sno\\' CO \'c r , uitable for ope ra ti ona l a\'a la nchc forecas tin g . J. Glaciol.. 35 ( 12 1 . 333- 342 . Brun . E" P. D avi d , \1. Sudu l a nd G. Brun o l. 1992. A numerical mode l lO simul a te sn ow-eOl'e r stral ig r ap h ) for operat io n a l a\'a la n che fo recast in g. J. Glaciol .. 38 ( 128 ). 13 22. Busc r , 0. , P. Fo hn, \\' . Good, H. Gublcr and B. Sa lm. 1985 . Diffe rent method s fo r th e a ssess mclll of ava la nc he dan ge r. Cold Reg . Sci. Tee/mol., 10 (3), 199-218. Bu se r , 0. , ~1. BLid er a nd \ \'. Good. 1987. Av a la nche fo recas t b y neares t neigh bour m e thod. II/ternllliol/ol Associolioll oJ I-/ )'drologicol Scieuces Publi({lliol/ 162 (Symposium a t D m'os 1986 - .J l'lIlal/che Forlllalioll . .llol'elllenl ol/d E;fJecls ), 557- 569. D U DE:"I I nformatik. 1988. EI/ C)'clojJaedia. ~I annheim , etc .. Dud ell\·e rl ag. Durand. Y .. E . Brun, L. ;\Ierind ol, G. Guyo marc·h. B. Lesa ffr e a nd E. ~ [ a rtin . 1993. A meteoro logica l estimation of rele\'a lll parameters for sno\\' models. AIIII . Glaciol .. 18, 65- 7 1. Fo hn , P. 1985. Das sc h\\'C izerische Law in e nbull et in - ei ne Illlerpre tatio n- shilfc fUr den BenLilzc r. Eidg. 1101. Se/lllee- ulld LIlIt'il1ell/orsch. '\I ill. 38. Fo hn , P. a nd P . H aec hl er. 1978. Prcv ision d e grosses a\'ala n eh es a u moye n d\1Il m od ele d etermini s tc-st a ti stiqu e. III Deurieme Rel/conlre IlIlernatiol/ole sur la N eige elles Avalanches . 12- 13 et /-Ianil 1978, Crenoble. Fral/ce. COIII/lles Relldus. Grenob le, Assoc ia ti on :"I a ti o na le pour I'Etud e d e la Nei ge et d es Ava la nches, 15 1 165. Fo hn , P. \\. B. a nd ]. ScJlII'cizer. 1995, \ 'e rifi cation of a\'ala n c he d anger wit h res pec t to ""a lan che fo recasting. III Sil'a rdi ere , F. , ed. The cOlllribulioll oj.lcielllijic resellrch 10 sllle!)' u:ilh SIlOIt' . ice olld amlallche . Atles de colloque. C/1II1110Ili.1 30 lIIai 3 juill 1995. Gre noble. Assoc iati on :"Iationale po ur J'Etud e de la :\' eigc er dcs ,\ \'a la nches AXE~A. 151 156. Fo hn. P .. W. Good. P. Bo is and C. Obled. 1977. E\'a luation and co mp a ri so n of stat isti ca l a nd cO I1\'C llli o n a l method s of fo recas ting ava lanche haza rd. J. Glaciol .. 19 f8 11, 375-387. G ira ud , G. 1991. ~IEPRA: m od clc ex pert d'aid e a la prh'ision du ri sque d 'a \·a lanc hes . III .. Icle.; du D'IIl/JOsilllll . . 1.\E,\ ·. I-CIS. /- I/t·.J/? , 4- 8 jn ill 19.9 1, Challlollir . Frallce. G re n oble, l\ ssoc ia ti o n :"I a ti onale po ur l' Etucle d e la N eige et d es i\\·ala n c hes . 248 254. Gi r a ud , G. , j. Lafe uill e and E. Pahaul. 1987. E\'aluation d e la Cjua li tc d e la p rc\'ision du ri sq u e d '