Design And Implementation Of Defense Model Of Tower Defense Game Based On Deep Reinforcement Learning

In the modern air defense system,the battlefield information is complex and rapidly changing.The use of AI for dynamic decision-making assistance has become an urgent need.Traditional AI models that rely on expert rules rely too much on rules and expert knowledge,and are not flexible enough.AI trained by deep reinforcement learning can learn from simulation or real battlefield data,making decision-making more flexible and reliable.Game Environment have always been one of the most ideal experimental environments for AI research.After investigation,I found that the tower defense game has similarities with air defense in terms of goal and operation which means the air defense scene can be simulated through a tower defense game engine.Therefore,this paper simulates the battlefield through the tower defense game engine and uses deep reinforcement learning to train the AI model.The main research work of this paper is as follows:Firstly I implement a tower defense game engine Mini-TD based on the RTS game engine provided by the open source deep reinforcement learning framework ELF.In order to make the game engine meet the experimental requirements and achieve the goal of simulating air defense combat scenes,the underlying logic is largely modified.The engine can meet the requirements of the experiment and is extensible.Secondly,I design a defense model of tower defense game based on deep reinforcement learning with this game engine,which uses game state data as input and outputs control commands.The LSTM neural network is used to deal with the time series problem.In order to select the object more accurately,the attention mechanism and Unit Mask are introduced,the A3 C algorithm is used to train the model.The model has achieved the expected goals in terms of model convergence speed,winning rate and performance evaluation.It can show that the model proposed in this paper is feasible and effective.It has the ability to provide dynamic auxiliary decision-making for modern air defense and the value of further research.