Hierarchical Bayesian Attractor Model for Dynamic Task Allocation in Edge-Cloud Computing

Edge computing responsive applications have been gaining attention in recent years, especially in AI technology. AI distillation techniques allow compact models to be placed at the edge or terminal, where computational power is limited, with lower latency than can be processed by AI in the cloud. However, AI with smaller models is generally less accurate, so the tradeoff between accuracy and latency, as well as power consumption, must be taken into account when determining processing assignments. Conventionally, such task allocation problems have required heuristic release due to computational difficulties. On the other hand, the heuristic release method has a problem of deviation from the optimal solution when the environment is quasi-static. Our research group takes the approach of continuously searching for the optimal solution in a quasi-static environment while immediately determining the quasi-optimal solution for dynamic environmental changes based on similarities with the past quasi-static environment. In particular, the Bayesian attractor model (BAM), which models brain decision making, is used to select a quasi-optimal solution based on similarity. However, BAM has the problem that appropriate selection becomes difficult when the number of alternatives increases. In this paper, we extend the BAM to a hierarchical model, inspired by the fact that in human decision making, related concepts and operations are grouped into chunks and organized hierarchically. We show that this allows us to maintain a high rate of correct responses even when the number of choices increases. We also investigate the composition of appropriate conceptual hierarchies and updating methods in the temporal hierarchy.