Affinity-Based Reinforcement Learning:  A New Paradigm for Interpretability

"The goal of this research project is to find ways in which to imprint reinforcement learning (RL) with preferred strategies that are interpretable. Agent interpretability is imperative for trust, deployment, and acceptance of future systems powered by artificial intelligence." Charl Maree says. Maree is an industrial PhD Research Fellow at Center for Artificial Intelligence at the University of Agder (UiA). The new paradigm – called affinity-based RL (ab-RL) - uses a fuzzy superposition of prototype RL agents which have a global affinity for taking or not taking certain actions. The composition of such prototypical agents is achieved thru hierarchical RL for specific agents with specific traits without the need for retraining of either the prototypes or the agents. Unlike other methods for controlling the search for optimum RL strategies which often need to alter the reward functions, our parsimonious method only uses regularization which guarantees convergence to an optimal strategy. Maree’s latest work which is an application to personalized financial advising which has been published in the prestigious journal Digital Finance (Springer) which demonstrates not only its novelty, but also value in real-world scenarios. Work is continuing to extend the paradigm to include local state- and time-dependent affinities with applications to the implementation of machine ethics, the investigation of climate change scenarios and interventions, the development of learning and teaching environments guided by student personalities, the optimal control of windfarms for remaining useful life extension, and the design of active matter. Future work may include this paradigm in multi-agent environments.