170 likes | 277 Views
Advanced Topics in Evolutionary Algorithms. Dan Shmidt Itay BiTtan. Outline. Operators Review Time and Fitness, Fitness and Time Results and interesting experiments Conclusions. Operators. Selection: Tournament selection of 4 Tournament selection of 2 Elite:
E N D
Advanced Topics in Evolutionary Algorithms Dan Shmidt ItayBiTtan
Outline Operators Review Time and Fitness, Fitness and Time Results and interesting experiments Conclusions
Operators • Selection: • Tournament selection of 4 • Tournament selection of 2 • Elite: • Decided to use elite to avoid loosing good solution along the way. • 10 percent of the population is Elite.
Operators • Cross Over: • Normal Cross Over: • Random Cross Over:
Operators • Mutation: • In some probability change parameter (i):Solution[i] = 1 – Solution[i] • Switching two parameters of same instance: Solution[i] = Solution[j] , Solution[j] = Solution[i] • Randomize first k parameters where k is randomSolution[0]=random, … , Solution[k-1]=random
Outline Operators Review Time and Fitness, Fitness and Time Results and interesting experiments Conclusions
Time Issues • Time is Allocated dynamically depending on: • Generation number. • Amount of individuals that are able to finish a lap. • Formula: Divided to three stages • Stage 1: before generation - allocated time is a constant • Stage 2: Every generation gets another ticks • Stage 3: If the number of individuals that finish one lap is greater then
Fitness Calculation • Fitness is the average of: • Why divide by time? – because time is dynamic and we want to be able to compare individuals who ran different amount of time. • Why average? – because we are keeping track of all of individuals runs and we don’t want to lose this information
Outline Operators Review Time and Fitness, Fitness and Time Results and interesting experiments Conclusions
Experiments First, we run the algorithm 19 times with different tracks. We got 19 sets of 22 parameters (each set evolved from a unique track). For each set we got a fitness function value (calculated during the process). Then, we run the each set on his corresponding track for a 10000 tics (which is enough to complete at least one lap) and we measured the real fitness value.
Experiments Details: elite, selection:4-tournament, population:30
Experiments Population 30 got better results (12 / 7)
Experiments Our algorithm also win 14 of 19!
Experiments Our algorithm win 14 of 19!
Experiments E-track-4 set win 14 of 19!
Outline Operators Review Time and Fitness, Fitness and Time Results and interesting experiments Conclusions
Conclusions A good solution to one track is a good solution on other tracks. The algorithm’s efficiency depends on the track. Our algorithm beats the original algorithm on most tracks.