Multi-Agent Reinforcement Learning System Using Value-Decomposition Network Algorithm in StarCraft Environment

Multi-Agent Reinforcement Learning (MARL) has been shown to be extremely successful in cooperative assignments. MARL allows for the control of multiple agents to complete multiple tasks in a certain environment and provide helpful services. In this paper, we will examine a war scenario with the StarCraft II Multi-Agent Challenges (SMAC) environment to implement a multi-agent system. For training, we employed one of the most popular MARL algorithms, which is the Value-Decomposition Network (VDN). This algorithm works on controlling the agents to cooperate with each other to achieve the desired goals. We will then use the battle won mean and the dead allies mean metrics to measure the performance of the VDN algorithm. The result showed that the VDN algorithm reaching the highest value of battle won mean with one million iterations and the lowest value of dead allies mean metrics with less than one million iterations. The hardware that we use in this work is CPU Cor i7 11800H, with 32 GB Ram and RTX 3080 laptop GPU, with CUDA Toolkit 11.7.1, and Pytorch 1.7.1.