Case Adaptation Using an Incomplete Causal Model 指導老師 ： 何正信教授 學生：潘立偉 學號： M 日期： 88/7/20 John D.Hastings, L. Karl Branting, and Jeffrey A. Lockwood, ICCBR, 1995 “Integrating Cases and Models for Prediction in Biological Systems”, L. Karl Branting, John D. Hastings & Jeffrey A. Lockwood, AI Application, under review “An Empirical Evaluation of Model-Based Case Matching and Adaptation”, L. Karl Branting, John D. Hastings, AAAI-94, 1994.

1999/7/20Li-we Pan2 What & Why in CARMA CARMA (CAse-based Range Management Adviser) integrate CBR with MBR to predict the behavior of biological systems characterized both by incomplete models and insufficient empirical data for accurate induction determining the most cost-effective response to a given pest infestation requires prediction crop or forage loss under each available option

1999/7/20Li-we Pan3 Cont. use of model-based adaptation as a technique for integrating CBR with MBR in domains in which neither technique is individually sufficient for accurate prediction. Under this approach, Case-based reasoning is used to find an approximate solution into a more precise solution

1999/7/20Li-we Pan4 Process description Using RBR to infer the relevant facts of the infestation case. Determine whether grasshopper consumption will lead to competition with livestock for available forage –Estimate the proportion of available forage that will be consumed by grasshoppers using CBR & MBR –Total the forage loss estimates for each subcase to predict the overall proportion of available forage that will be consumed by grasshoppers

1999/7/20Li-we Pan5 Cont. –Compare grasshopper consumption with the proportion of available forager needed by livestock If competition, determine what possible treatment options should be excluded using rules If there are possible treatment options, for each one provided an economic analysis by estimating both the first-year and long-term savings using rule-based, model-based, and probabilistic reasoning

1999/7/20Li-we Pan6 Prototypical cases are not expressed in terms of observable features, but rather in terms of abstract derived features are extended in time, representing the history of a particular grasshopper population over its lifespan

1999/7/20Li-we Pan7 CARMA Step Determining relevant case features case matching model-based adaptation –(x)case factoring –temporal projection –feature adaptation –critical-period adjustment forage loss estimation determining treatment options treatment recommendation

1999/7/20Li-we Pan8 Model-based reasoning Case factoring –spilt the overall population into subcases of grasshopper with distinct overwintering types Temporal projection –retrieval all prototypical cases whose overwintering types match that of the subcase. –CARMA must project the best matching prototypical case forward or backwards in time to align its average developmental phase with that of the new subcase –CARMA breaks the distribution into daily populations, projects the populations the required number of days

1999/7/20Li-we Pan9 Featural adaptation Modify to account for any featural differences between it and the subcase. FL(NC) = FL(PC) +  Ai * QFD(i) QFD = (Q(NC, i) - Q(PC, i)) / Q(PC, i) A : adaptation weights, QFD : quantitative difference for feature I between the new case and prototypical case

1999/7/20Li-we Pan10 Critical Period Adaptation Grasshopper consumption is most damaging if it occurs during the critical forage growing period must be adapted if the proportion of the lifespan of the grasshoppers overlapping the the critical period in the new case differs form that in the prototypical case Ex: subcaseA : 47% case8 : 6% (47-6)/6 = * adapt weight

1999/7/20Li-we Pan11 Learning match and adaptation Weight Match weights (by system) –determining the mutual information gain between case feature and qualitative consumption categories in a given set of training cases Featural adaptation –use hill-climbing algorithm –to min RMSE for prototypical case library P and match weights M, PFL : CARMA’s predocted forage loss, ExpertPred : expert’s prediction of consumption for each training cases Ci

1999/7/20Li-we Pan12 Algorithm Function AdaptWeights(t, p, M) I <= initial increment D min <= minimum improvement threshold I min <= minimum increment threshold A <= initial list of global adaptation weights D’ <= RMSE(T, P, M, A) D <=  loop until (I < I min ) do loop until (|D’-D| < D min ) do D <= D’  <= the change to an element of A by I for which RMSE(T, P, M,  (A)) is least D’ <= RMSE(T, P, M,  (A)) if (D’ < D) then A <=  (A)else D’ <= D I <= I/2 return A

1999/7/20Li-we Pan13 Test function Fuction LeaveOneOutSpecificTest(T) for each case Ci  T do P := T - Ci;prototypical cases M := global match weights for set P according to info. Gain for each prototypical case Pj  P do T := P - Pj; training set Pj(A) := Adaptweights(T, {Pj}, M) Di := (PredictForageLoss(Ci, P, M) - ExpertPred(Ci))^2 return((Avg(D)^(1/2)) Fuction LeaveOneOutGloubleTest(T) for each case Ci  T do P := T - Ci;prototypical cases M := global match weights for set P according to info. Gain G := Adaptweights(T, {Pj}, M) Di := (PredictForageLoss(Ci, P, M, G) - ExpertPred(Ci))^2 return((Avg(D)^(1/2))

1999/7/20Li-we Pan14 Schedule