Learning Classifier Systems Learning Classifier Systems (LCS) • The system has three layers: – A performance system that interacts with environment, – An apportionment of credit algorithm that rates rules as to usefulness, – A rule discovery algorithm that generates plausible new rules to replace less useful rules. Performance System Cycles • Message is posted in the message list from the input interface. • Each rule is matched against the message list • All matching rules compete to post in the next message list via bidding process; winning rule posts in the new message list • The output interface checks the new message and produces an effector action. • The new message list replaces the previous one. • Repeat. Overview of LCS Rule format • Rule – Condition = {0,1,#}k – Action = message to be posted in the message list – Strength = rule’s usefulness to the system Example (Wolf or Grandmother?) Encoding teeth kind ears num. of legs smart scream runaway kiss Wolf 1 0 1 1 0 1 1 # 1 1 0 1 # 0 0 1 GrandMa 0 0 Matching Message List 0 [N] [M] 1 0 0 Condition Action Strength # 1 # # #010 100 1 # 0 1 1### 50 0 # 0 # 0011 100 1 # # # 1010 1000 1 0 1 # 0111 1000 Action Strength Condition # 1 # # #010 100 0 # 0 # 0011 100 Bidding Process Bid(R,t) = β × specificity(R) × Strength(R,t) Specificity(R)= number of non # / k [M] Rule id Condition r1 # 1 # r3 0 # 0 Action Strength # #010 100 # 0011 100 β = 0.2 Bid(r1) = 0.2 × ¼ × 100 = 5 Bid(r3) = 0.2 × ½ × 100 = 10 r3 posts its message in the new message list. Credit assignment: Bucket Brigade r3 coupled r5 executed Environment Bucket 10 Bucket 150 Reward 200 Credit assignment: Bucket Brigade r3 Bucket 10 r5 Bucket 150 Environment Reward 200 Genetic Algorithms • Fitness = rule strength • Parents: Strong classifiers (best, roulette wheel, etc.) • Mutation: alter parts of parent’s string • Crossover: exchange parts of parents’ strings • Offspring replaces a weak rule. Genetic Algorithms (cont.) Crossover Parent 1 Parent 2 Crossover point 0 0 1 0 1 1 # # 0 0 1 0 1 1 0 0 1 0 1 0 0 1 0 0 1 0 1 0 0 1 # # Mutation Parent 1 Parent 2 0 0 1 0 1 1 # # 0 0 1 0 1 0 # # 1 0 1 0 0 1 0 0 1 0 1 0 0 1 0 0 Maze Environment (Signal smell-ahead bump heading score location) 40 5 f N 5 (1,2) A Environment Message List GF Condition Action Strengt h # >0 # # # # GF 1000 # <0 # # # # ∧ TL TL 1000 # <0 # # # # ∧ TR TR 1000 References • A Mathematical framework for Studying Learning in Classifier Systems, John H. Holland, Phsyca D, Vol 2, No 1-3, 1986, pp. 307-317 • A First Order Logic Classifier System, Drew Mellor Gecco ’05