Evolving robots to play dodgeball
Daniel and Uriel
Creative Commons CC BY 4.0
In nearly all videogames, creating smart and complex artificial agents helps ensure an enjoyable and challenging player experience. Using a dodgeball-inspired simulation, we attempt to train a population of robots to develop effective individual strategies against hard-coded opponents. Every evolving robot is controlled by a feedforward artificial neural network, and has a fitness function based on its hits and deaths. We evolved the robots using both standard and real-time NEAT against several teams. We hypothesized that interesting strategies would develop using both evolutionary algorithms, and fitness would increase in each trial. Initial experiments using rtNEAT did not increase fitness substantially, and after several thousand time steps the robots still exhibited mostly random movement. One exception was a defensive strategy against randomly moving enemies where individuals would specifically avoid the area near the center line. Subsequent experiments using the NEAT algorithm were more successful both visually and quantitatively: average fitness improved, and complex tactics appeared to develop in some trials, such as hiding behind the obstacle. Further research could improve our rtNEAT algorithm to match the relative effectiveness of NEAT, or use competitive coevolution to remove the need for hard-coded opponents.