A Surprise Triggered Adaptive and Reactive (STAR) Framework for Online Adaptation in Non-stationary Environments
DOI:
https://doi.org/10.1609/aiide.v5i1.12364Keywords:
online adaptationAbstract
We consider the task of developing an adaptive autonomous agent that can interact with non-stationary environments. Traditional learning approaches such as Reinforcement Learning assume stationary characteristics over the course of the problem, and are therefore unable to learn the dynamically changing settings correctly. We introduce a novel adaptive framework that can detect dynamic changes due to non-stationary elements. The Surprise Triggered Adaptive and Reactive (STAR) framework is inspired by human adaptability in dealing with daily life changes. An agent adopting the STAR framework consists primarily of two components, Adapter and Reactor. The Reactor chooses suitable actions based on predictions made by a model of the environment. The Adapter observes the amount of “surprisingness” and triggers the generation of new models accordingly. Preliminary experimental results show that STAR agents are competitive in performance as compared with current approaches, while being much more cost-effective by avoiding the negative effects of historical data. Furthermore, since response and adaptability are decoupled in the framework, the adaptive component can benefit other autonomous agents in a variety of domains with non-stationary environments.