Adaptive Memory Module for Sequential Planning and Reasoning

Kshitij Gupta, Sean Spinney, Xutong Zhao, Janarthanan Rajendran, Patricia Conrod, Irina Rish, Sarath Chandar

Keywords: Adaptive computation, Memory, Planning, Reasoning, Offline RL

Abstract: Efficient planning and reasoning in sequential decision-making tasks remains a core challenge for machine learning models. These tasks often involve intricate decision sequences leading to combinatorial complexity that hampers traditional planning methods. Humans on the other hand leverage flexible planning strategies and adapt their thinking time based on the complexity of the problem at hand to efficiently solve complex reasoning problems. Inspired by this, we propose and investigate an end-to-end memory-based adaptive learning algorithm to enhance planning capabilities and resource allocation of AI agents. Our study borrows concepts from adaptive computation and incorporates memory and reusability mechanisms into agents. This allows agents to meta-learn flexible reasoning strategies, plan deeper, and efficiently adjust their computation to not only improve inference time efficiency but also generalize to more complex problems. Finally, our study of the adaptive memory module reveals patterns comparable to human decision-making mechanisms such as increasing certainty, reconsideration and alternative exploration. This work contributes to the evolving understanding of harnessing adaptive computation to enhance machine learning models' capabilities in complex reasoning and sequential decision-making tasks.

Primary Area: applications to robotics, autonomy, planning

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/2024/AuthorGuide.

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: 8794