M. Gams Jozef Stefan Institute BASIC AI METHODS M. Gams Jozef Stefan Institute
PURPOSE: Show how AI methods work - to con-computer specialists - with a simple, understandable example
THE EXAMPLE: Solution: 1 2 3 Possibilities 1 1 1 2 2 2 3 3 3 3 x 3 x 3 = 27 possibilities 100 3 … no. of all particles in the universe
GENETIC ALGORITHMS Population evolution: breeding, selection; competition, better adapted offsprings – only the best survive Genetic code of an individual is represented by a sequence of digits, 3 seeds: 111, 222, 321 Cross-over (breeding) 111 + 222 122(2), 212(0), 221(1), 211(0), 121(2), 112(1) - () 123 Mutation (each third individual, next position, +1) 222(1), 212(0), 221(1), 221(1), 121(2), 112(1)
GENETIC ALGORITHMS 111, 222, 321 1+2 222(1), 212(0), 221(1), 221(1), 121(2), 112(1) 1+3 122(2), 311(0), 321(1), 111(1), 121(2), 111(1) 2+3 231(0), 321(1), 322(1), 323(2), 222(1), 221(1) 3 BEST: 121(2), 122(2), 323(2) NEXT STEP: SEVERAL SOLUTIONS 123(3)
RULES: if high_fever then illness if fever > 37 then illness if (axilliar = yes) and (degree of diff = fairly) and (lung = no) and (sex = female) then breast (100%) Solution: if x1=1 and x2=2 in x3=3 then (3).
RULES: Start: 111 2 candidates First step: if x1=1 then (1) if x1=2 then (0) if x1=3 then (0) Next step: if x1=1 and x2=2 then (2) if x1=2 and x2=2 then (1) Next step: if x1=1 and x2=2 and x3=3 then solution (3)
Idea – repeat splitting the space of all possibilities 11(1) 12(2) 13(1) 121(2) 122(2) 123(3) x1=1 2x2x3 3x2x3 1x2x3 NE DA x2=2 x3=3 123 11x3 13x3 12x3 TREES: Idea – repeat splitting the space of all possibilities
Output(neuron) = 1 if Σ wi xi > C NEURAL NETWORKS: Output(neuron) = 1 if Σ wi xi > C 0 otherwise Our case: 9 connections if w1 = 1, w5 = 1 w9 = 1, Σ wi xi = 3 100010001 123123123 x1 x2 x3
NEURAL NETWORKS: Learn weights wi examples in a sequence 3 3 3 (1) 111111111 (3) 3 3 2 (0) 110110111 (1) 3 3 1 (0) 110110101 (1) 3 2 3 (2) 110110001 (2) 3 2 2 (1) 110110001 (1) 3 2 1 (1) 110110001 (1) 3 1 3 (1) 110110001 (2) 3 1 2 (0) 110010001 (0) 3 1 1 (0) 110010001 (0) 2 3 3 (1) 110010001 (2) 100010001
CONCLUSION - DIFFERENT AI SEARCH METHODS GENETIC/EVOLUTIONARY METHODS RULE-CONSTRUCTING MACHINE LEARNING TREE-CONSTRUCTING NEURAL NETWORKS - AROUND 10 ADDITIONAL METHODS SOME DETERMINISTIC, OTHER SOFT - DIFFERENT METHODS APPROPRIATE FOR DIFFERENT PROBLEMS AND TASKS - EFFECTIVE, MANY TOOLS http://ai.ijs.si/