A comparative study of NEAT and XCS in Robocode

Historically, learning algorithms have been applied to games as a test of their performance, and with the exponential increases in available computational power, machine learning has been attempted in increasingly complex environments. This paper details the application of neuroevolution of augmenting topologies (NEAT) and accuracy-based learning classifier system (XCS) to the Robocode game environment, with the aim of discovering the ability of each algorithm to learn in this environment. Existing implementations of each algorithm were modified and augmented to allow them to operate in Robocode. In the experiments, performance is measured by pitting the algorithmically-driven players against a series of opponents in various tactical challenges. We conclude by discussing the comparative advantages and disadvantages of both NEAT and XCS as applied to Robocode. Both are able to learn competent strategies, but NEAT is susceptible to overfitting and XCS struggles when many actions are available. The selection of appropriate training scenarios is shown to be a key factor in ensuring that evolved strategies are maximally general for both algorithms.