Neuroevolution by Particle Swarm Optimization with Adaptive Input Selection for Controlling Platform-Game Agent
Akira Hara, Jun-ichi Kushida, Koji Kitao, Tetsuyuki Takahama · 2013
Neuroevolution has been widely used for action control of agents. Agent controllers are represented by Neural Networks (NN), and the connection weights and/or the structure of NN are optimized by evolutionary computation such as Particle Swarm Optimization (PSO). The agent's perceptual inputs are used as the inputs of NN. When the framework is applied to the agent control in platform games where a lot of perceptual information is available, the number of nodes in the input layer becomes enormous if all the information is used. Therefore, only the necessary information should be selected and used as the NN inputs, but it is difficult to select the appropriate information beforehand. In this research, we propose a new PSO method which can optimize not only the connecting weights of NN but also the selection of the perceptual information simultaneously. By our method, the increase of network size can be prevented and the controllers can be optimized efficiently. We examined the effectiveness of our method in the Mario AI Championship.