Applying and Comparing Evolutionary Algorithms for Robot Tanks

Robocode, an open source tank combat game, has become immensely popular based on both its educational as well as its “fun " value. Throughout the years, many interesting AIs have been developed for Robocode, most of which were based on finite-state machines and were generally “hard-coded". The effects of employing evolutionary algorithms and machine learning on Robocode, however, have not been so thoroughly explored. In his paper, we focus on two main methods and their effectiveness in evolving highly-intelligent robots which can effectively defeat good, existing hard-coded robots. A traditional Genetic Programming (GP) approach is first used as an intermediate goal due to its relative-ease as well as in order to familiarize ourselves with using machine learning on Robocode. We then employ Neuroevolution of Augmenting Topologies (NEAT), a method of evolving complex, unfixed topologies in Neural Networks since it has historically achieved success in evolving highly complex behavior. Although our results after experimentation were not we hoped, the trend in data leads us to believe that further modification of our experimental method can possibly help us attain the perfect battle bot