Hybrid Learning Algorithm In Real-time Strategy Games

In a typical attack/defense scenario of real-time strategy (RTS) games, we present ahybrid evolutionary algorithm optimization model to handle the optimization problem ofmulti-team weapon target assignment, and develop a speed up model to meet the real-timerequirement of games. The research work is mainly divided into two parts.Firstly, we use genetic algorithm (GA) to assign the number of different types ofweapons under limited resources. Then we put the obtained optimal results into a randomgame’s map to obtain final defensive locations. In this process, GA and particle swarmoptimization (PSO) are used to achieve the best distribution of defensive positionsrespectively and their performance is compared in RTS Games.Secondly, we propose a two-stage speed up model using the combination of geneticalgorithm and ant algorithm (GA-AA) and then artificial neural network(ANN) to enhance theefficiency and adaptive ability of the previous model. In the first stage, we use GA-AA tospeed up the off-line training process. GA is used in initial training, and AA takes place of GAwhen its “pheromone” is sufficient. The off-line results, i.e., the optimal positions ofdefensive weapons are finally obtained. In the second stage, an ANN is further combined inthe model to accelerate the online process. The off-line results with neighbor information areinput the ANN as training cases. The finally trained ANN can quickly estimate the weapondefensive distribution in randomly generated maps. Compared with the traditionalevolutionary algorithms, the two-stage speed up model has improved the efficiency of bothoff-line and on-line training processes and has better adaptive ability to random game maps.