Genetic Algorithm

Code runs in python 2 To improve performance of Pacman agents using GA - The fitness is determined by testing on 100 randomly generated scenarios where the max cycle per scenario is 5000, reward strategies that are at least 3 blocks close to prey (the closer the better), the mutation rate is set at 0.01, 2 point crossover with a crossover probability of 0.4, a tournament selection scheme with a tournament size tour of 5 is used to select the parents at each generation. The following pseudocode describes the methodology: 1. Repeat the following for 10 trials on selected prey: (a) Randomly generate a population of 100 individuals. (b) Repeat the following until the predators show no improvement after 200 generations: i. Simulate each predator strategy on 100 scenarios and evaluate its fitness based on the performance on those scenarios. ii. Select 100 individuals from the current population using Tournament selection, pair them up, and create a new population by using 2-point crossover with mutation. (c) The best strategy found over all generations is used as the solution of this trial. The fitness of this strategy is then recomputed by testing on 1000 new randomly generated scenarios. 2. The strategy that performed best over all 10 trials is used as the solution to this prey.