Self-Improving Logic from Experimental Observations
Keywords: self-improving logic, reinforcement learning, hierarchical learning, second-order logic, markov decision process, group representations, tokenisation, model-based learning, dynamics models, machine understanding, unsupervised learning, computational learning theory, non-convex optimisation, wahba's problem, skill-based reinforcement learning, skills pruning, transfer learning
TL;DR: Learned representations of actions in topological groups as an humanly-understandable alternative to neural networks to efficiently structure a model-based autonomous agent's dynamics prediction and planning within a Markov Decision Process.
Abstract: Learning relevant, transferable representations of actions to drive model-based reinforcement learning processes stands as a major challenge in robotics on the path to general-purpose autonomous agents, equivalent in their reasoning power to Large Language Models. To this end, we introduce a novel framework which allows autonomous agents to learn how to represent their actions as high-dimensional rotations over the system’s observations. We then show how such representations may be considered optimal under the assumption that actions are distance-preserving, and present how these representations of low-level actions can be composed to represent sequences of actions and allow for multi-scale hierarchical learning and long-horizon planning. We finally discuss schemes to compare such representations in order to allow for a better informed transfer of skills across tasks and better understand the agent’s behaviour, before conducting experiments using a modified TD-MPC2 agent to better quantify $in$ $concreto$ the limitations of our framework.
Primary Area: interpretability and explainable AI
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 12021
Loading