Using NEAT to Evolve Cooperation and Intelligence in Game Theory Problems

The goal of our study was to see if neural networks would cooperate with each other in two different game theory problems. We replicated a study entitled “Cooperation and the Evolution of Intelligence” by Luke McNally, Sam P. Brown and Andrew L. Jackson, however we evolved the neural networks with NEAT instead of a genetic algorithm. The neural networks would play every other individual in the population in the two game theory problems, choosing to either cooperate or defect and receiving a payoff based on their action and the action of their opponent. We evolved 10 runs of each problem to 10,000 generations, measuring the frequency of cooperation in the each generation. One of the game theory problems, known as the Iterative Snowdrift Dilemma, by the virtue of its structure, encouraged cooperation more than the other game theory problem, the Iterative Prisoners’ Dilemma. Our hypothesis was the neural networks evolved with the Snowdrift Dilemma would have higher levels of cooperation than neural networks evolved with the Prisoners’ Dilemma. Our results supported the hypothesis and led to some insights into the behavioral strategies of the neural networks and the different number of neurons in the networks in the different problems.