The Design And Research Of Multi - Agent Model In Real - Time Strategy Game

Artificial Intelligence (AI) since it’s originated, with the development of theory and technology, has been gradually used in various fields. The Artificial Neural Network(ANN) is one of the most important branches in AI, from the traditional neural network that is designed by developer to the evolutionary neural network (ENN) that is combined by evolutionary algorithm (EA) an ANN. The technology based on ANN has used has solved many problems in various fields like, auto-control, pattern recognition, multi-agent system and so on.Although RTS (Real time strategy game) is a branch of computer games, because it contains many characteristics in the field of AI, such as decision making under uncertainty, large state and action spaces, multi-agent cooperation and so on. In addition, RTS provides a complex and realistic simulation of environment for AI. The simulation provides intuitive and real-time feedback and few of the messy properties and cost of hardware on the realistic environment that make it as a perfect platform for the research of AI.As a game that provides a simulation of the real world war, there will be multiple agents in the scene. The problems that how to control these multiple agents effectively and make them cooperate with each other to complete the target are the main content of this paper. In this paper, a multi-agent model designed by ENN can cooperate with other agents to complete the target with high efficiency. Therefore, the main research work and achievements include the following points:(1) the conclusion of research that using RTS as platform for the research of AI such as Reinforcement Learning, Case-Based Reasoning, Game-Tree Search and Monte Carlo Planning and so on; (2) Summary the characteristics of ENN. Use the Fixed Topologies Evolutionary Neural Network to design the multi-agent model. Combine the BP and GA to apply to the multi-agent model, so as to enhance the performance of Multi-Agent in a test environment; (3) Improve the general Point Crossover operator and call it Border Point Crossover operator. The performance of ENN based on GA which using the Border Point Crossover is improved at same level; (4)Apply a new ENN, namely Neuro Evolution of Augment Topologies (NEAT) to design the multi-agent model. The whole NN that controls multi-agent based on NEAT. This paper will prove the effectiveness of this multi-agent model based on NEAT though experiment.