Adaptive Probabilistic Learning by Collectives in Dynamic Environments

Probability Collective (PC) is a methodology for distributed optimization by sampling an explicitly parameterized probability distribution over the space of solutions. This parameterization effectively utilizes granules of probability distributions to construct computational models for solving complex systems-level optimization problems. In this paper we present a study of using this probabilistic collective learning framework for adaptive optimization in the context of dynamic environments. Two scenarios of PC in dynamic optimization tasks are compared: PC₁ (original PC settings), PC₂ (the probability distributions are reset to uniform when an environment changes). By allowing PC to re-explore the search space, we show that PC₂ is more adaptive to environmental changes, thereby outperforming the original PC in rate of descent as well as long term extrema-tracking optimization. The study of the PC in changing environments therefore sheds light into how this probabilistic learning methodology utilizes the features of granular computation to solve complex dynamic optimization problems.