Evolving Team-Agent Based On Concurrent Evolutionary Artificial Neural Networks

Evolutionary Artificial Neural Networks (EANNs) has been highly effective in Artificial Intelligence (AI) and in training Non-Playable-Characters (NPCs) in video games. When EANNs is applied to design game NPCs' smart AI which can make the game more interesting, there always comes two inportant problems:(1) the more complex situation NPCs are in, the more complex structure of neural networks needed which leads to large operation cost;(2) how to design adaptive fitness functions that can make the proper evolution.In this dissertation, the Concurrent Evolutionary Neural Networks (CENNs) isproposed based on EANNs which deletes or fixes the connection of the neurons to reduce the operation cost in evolution and evaluation process. Darwin Platform is chosen as our test bed to show its efficiency: Darwin offers the competitive team game playing behaviors by teams of virtual football game players. The Red team and the Blue team are competing in the soccer field, the field players in Red team are evolved during the virtual game playing. The Blue programmed by FSM (Finite State Machine) team leads the evolution successful.The experimental results show that the soccer filed players in Red team can find strategic position by training with CENNs and prove the efficiency of the algorithm.