A Model of Predicting Time Series for Multi-Agent Opponent Based on Gate Recurrent Neural Network
HU Zhiyao
YU Miao
TIAN Kaiyuan
Abstract：A large number of studies shows that multi-agent reinforcement learning algorithm is difficult to converge when the number of agents increases.However,centralized training can overcome the challeng-ing non-stationary environment under the condition of predicting the action of the opponent,achieving still better learning results.In the light of predicting opponent's continuous decision-making,the time series opponent modeling method based on gated recurrent neural network is proposed.Based on the multi-agent reinforcement learning framework MADDPG,the opponent's state,environmental state,and our agent's action are utilized for constructing time series to serve as the input of gated recurrent units,aiming to cap-ture the time series information of opponent agents.A length-variable model inference method is presented to determine the optimal length of the input time series,avoiding the misuse of stale training samples in-curred by the adjustment of the opponent's strategies.The simulation results show that the proposed op-ponent model can effectively predict the opponent's future actions in combination with the length-variable model inference method.Compared with the SAM method,an improvement on the prediction accuracy is over 6％ by this method.
Keywords：deep reinforcement learningmultiple agentsgated recurrent unitstime series predictionop-ponent modeling
Publication Date：2025-12-25
Online Publishing Date：2025-12-26(First online date of this platform, not the publication date of the document)
Pages：10( 106-115 )