Evolving Game Agents Based on Adaptive Constraint of Evolution

Neuro-evolution (NE) has been highly effective in artificial intelligence (AI). Evolving multi-weights neural network (MWNN), which is always used in training game agents, is a kind of NE system. An important question in evolving MWNN is how to code real values onto binary strings and how to escape from premature convergence. We suggest a method called adaptive constraint of evolution (ACE), which can solve both problems of evolving MWNN represented as an non-player-character (NPC) in games based on real-coded genetic algorithm (RCGAs). ACE is then evaluated in training game agents to show its efficiency.