Real-time strategy game micro for tactical training simulations
Complex, realistic scenarios in training simulations can benefit from good control of large numbers of simulation entities. However, training simulations typically focus on simulation physics and graphics over the intelligence required to control large numbers of entities. Real-Time Strategy games, on the other hand, have evolved to make tradeoffs between the AI needed and human interaction required to control hundreds of entities in complex tactical skirmishes. Borrowing from work in real-time strategy games, this paper attacks the problem of controlling groups of heterogenous entities in training simulations by using a genetic algorithm to evolve control algorithm parameters that maximize damage done and minimize damage received during skirmishes in a real-time strategy game-like simulation. Results show the emergence of complex, coordinated behavior among groups of simulated entities. Evolved behavior quality seems to be relatively independent of the underlying physics model but depends on the initial dispositions of entities in the simulation. We can get over this dependence and evolve more robust high performance behaviors by evaluating fitness in several different scenarios with different initial dispositions. We believe these preliminary results indicate the viability of our approach for generating robust, high performance behaviors for controlling swarms of entities in training simulations to enable more complex, realistic training scenarios.
Louis, Sushil J., Tianyi Jiang, and Siming Liu. "Real-time strategy game micro for tactical training simulations." In Proceedings of the Genetic and Evolutionary Computation Conference Companion, pp. 1656-1663. 2018.
GECCO 2018 Companion - Proceedings of the 2018 Genetic and Evolutionary Computation Conference Companion