GTLMA: Generalizable Hierarchical Learning for Tasks with Variable Entities

The current ALlocator-Actor Multi-Agent Architecture (ALMA) has shown convincing results on some complex composite tasks. The two-level decision-making process used by ALMA, along with a high-level subtask allocation strategy and low-level agent policies, is the secret to its success. Although ALMA’s modularity allows it to generalize to different environment settings to some level, we believe there is still opportunity for improvement in this area. Due to the intrinsic property of ALMA, the fixed number of entities in the model architecture for a specific task, it may even get stuck in scaling to similar task which only differs in the number of entities. Therefore, we introduce a new module called Knowledge Container, which has a significant effect on improving the performance and generalizability of agents. We demonstrate our method can obtain and keep the considerable performance without any re-training or fine-tuning.