Understanding and Controlling a Maze-Solving Policy Network

Ulisse Mini, peli grietzer, Mrinank Sharma, Austin Meek, Monte MacDiarmid, Alexander Matt Turner

Primary Area: societal considerations including fairness, safety, privacy

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.

Keywords: interpretability, alignment, AI safety, reinforcement learning

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: Deep convolutional policy network contains redundant and manually-controllable internal representations of goal location.

Abstract: To understand the goals and goal representations of AI systems, we carefully study a pretrained reinforcement learning policy that solves mazes by navigating to a range of target squares. We find this network pursues multiple context-dependent goals, and we further identify circuits within the network that correspond to one of these goals. In particular, we identified eleven channels that track the location of the goal. By modifying these channels, either with hand-designed interventions or by combining forward passes, we can partially control the policy. Our work shows the goals of this network are redundant, distributed, and re-targetable, shedding light on the behaviour of AI systems and their goal-directed behaviour.

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: 1951