USING NEURAL NETWORKS TO DETERMINE COURSE OF ACTION FOR A LAND-BASED CONSTRUCTIVE SIMULATION

Abstract

The latest breakthroughs achieved by various commercial companies in applying neural networks to games have led to research in applying such technologies to the military domain. This thesis uses a proxy stochastic simulation environment to identify the factors required to train a neural network via reinforcement learning to maneuver forces in a battlefield and overcome hostile forces. An incremental training approach was used to train the neural network to recognize the enemy first, before training it to consolidate forces as a form of emergent behavior. Eight neural networks were trained successfully to maneuver forces to engage the enemies. Seven of these neural networks managed to generalize their training to search for enemies in a larger area than the scenario in which they were trained. The neural networks also showed some success in performing force consolidation before attacking an enemy. The most important neural network hyperparameters that contributed to training success were training duration, learning rate, discount factor, and loss factor. This thesis also found that the neural network reward function and randomizing starting positions during the training phase are critical to training success.