Efficient Active Action Selection with Causal Effect in Uncertain Environments

Abstract

In the actual environment, animals, including humans, can not observe environmental states directly because of their uncertainty. Active inference based on the free energy principle is a framework that attempts to know the environment by taking actions and intervening in such an uncertain situation to change it to a more predictable state. However, the environmental state is changed by not only the agent's actions but also latent factors (e.g., other agents and environmental characteristics). To solve this issue, we propose a new framework of active inference using causal effects that can infer environmental changes only caused by the agent to remove latent factors from state changes. Experimental results showed that the proposed framework outperformed on a decision-making task efficiently than a standard active perception framework.