Resource Allocation Problems With Reinforcement Learning

Sammanfattning

This thesis investigates using reinforcement learning to solve problems with resource allocation often present in strategy games. The specific problem has to do not only with allocating resources, but also distribution of the resources via restricted means. It tackles challenges around temporal and collaborative aspects in resource distribution. Four different reinforcement learning algorithms are used, tabular Q-learning, deep Q-learning, advantage actor-critic and proximal policy optimization. This variation includes both tabular and approximative, as well as value based and policy based methods. Three kinds of architectures are developed to use the methods on the problem, one for the tabular method, one global architecture and one local architecture. The results show that the tabular method can effectively deal with small problems, but run into issues with spatial complexity as the problem size increases. For the approximative methods, the global architecture shows great potential if given lots of training time when using advantage actor-critic or proximal policy optimization. The local architecture shows slightly worse potential but with less training time, using only deep Q-learning. Overall the results are promising for using reinforcement learning for problems in resource allocation in strategy games.