USING CONVOLUTION NEURAL NETWORKS TO DEVELOP ROBUST COMBAT BEHAVIORS THROUGH REINFORCEMENT LEARNING

Abstract

The success of reinforcement learning (RL), as shown with video games such as StarCraft and DOTA 2 achieving above-human performance levels, begs questions about the future role of the technology in military constructive simulations. The objective of this study was to use convolutional neural networks (CNN) to develop artificial intelligence (AI) agents capable of learning optimal behaviors in simple scenarios featuring multiple unit and terrain types. This thesis sought to incorporate a multi-agent training regimen that could be employed in the domain of military constructive simulations. Eight different scenarios, all with varying levels of complexity, were used to train agents capable of exhibiting multiple types of combat behaviors. Overall, the results demonstrate that the AI agents can learn robust tactical behaviors required to achieve optimal or near-optimal performance in each scenario. The findings additionally indicate that a better understanding of multi-agent training was attained. Ultimately, CNN combined with RL techniques prove to be an efficient and viable method to train intelligent agents in military constructive simulations, and their application can potentially save human resources in the execution of live exercises and missions. It is recommended that future work should investigate how to best incorporate similar deep-RL methods into an existing military program of record constructive simulation.