New Framework for Embedded Agency and Multi-Agent Learning

A novel mathematical framework focused on prospective learning and embedded agency has been developed, expanding upon the principles of universal artificial intelligence (AIXI). This framework tackles issues in multi-agent reinforcement learning, particularly where non-stationarity is caused by the learning processes of other agents. It emphasizes self-prediction, enabling Bayesian RL agents to foresee both their future perceptual inputs and actions. Traditional model-free reinforcement learning theories assume stationary dynamics and independent agents, which fall short in multi-agent contexts. The new framework integrates agents into the environment, recognizing that other agents hold beliefs about them. This research is accessible on arXiv with the identifier 2511.22226v2, categorized as a 'replace' announcement. The abstract provides insight into its motivation and key concepts, though the complete content is not included. This study is pertinent to artificial intelligence, machine learning, and multi-agent systems.