Enhancing AI Performance in Real-Time Strategy via StarCraft II Learning
Vikrant Shokeen, Amit Kumar Sharma, Sandeep Kumar, Ahmad Taher Azar, Nashwa Ahmad Kamal, Chakib Ben Njima · 2025
The outcomes of the StarCraft II Learning Environment (SC2LE) were set out to be recreated in order to replicate the accomplishments of Google DeepMind’s groundbreaking work for five different minigames. Several approaches were used to create a baseline, beginning with agents that followed pre-scripted sequences and performed random activities. Additionally, the effectiveness of Q-learning was assessed in this situation. The Asynchronous Advantage ActorCritic (A3C) training algorithm was subsequently explored. As a result of domain-specific knowledge covered by the A3C, both entire action and observation spaces were strategically reduced. The A3C was particularly exciting as it aimed to utilize the intrinsic qualities of the games to accelerate learning. Furthermore, experimentation was conducted with the MetaLearning Shared Hierarchies (MLSH) algorithm to develop a greater degree of transferable skills, as this algorithm intended to enhance strategy generalizability across the various minigames. Therefore, the dynamics of these minigames and the efficacy of alternative learning strategies are provided through careful experimentation and analysis in this paper.