Comparison of NEAT and HyperNEAT Performance on a Strategic Decision-Making Problem
Jessica Lowell, Sergey Grabkovsky, Kir Birger · 2011
Neuroevolution is a useful machine learning approach for problems with limited domain knowledge, but it has not done well with strategic decision-making problems, where the correct action varies sharply as the agent moves across states. Two promising neuroevolution algorithms are Neuro Evolution of Augmenting Topologies (NEAT) and its extension, Hyper NEAT. We compare the performance of these two algorithms on a benchmark problem, Keep away Soccer, that requires strategic decision-making. Our results demonstrate that Hyper NEAT outperforms NEAT on a simple instance of the problem but that its advantage disappears when the problem is complicated.