Previous work on preference learning focuses extensively on using rewards as proxies. Despite fitting into the reinforcement learning paradigm nicely, reward-based machine learning approaches face the difficulty of fully representing personal preferences as rewards and the challenge of updating the policy with few samples. In this study, we aim to take an alternative rule-centric approach – drawing inspiration from cognitive science and building a decision-making framework centered around production rules. The production rules in the cognitive framework are abstract, modular, interpretable, and composable - all important features in human-AI interactions. Therefore, we propose a framework that combines the general knowledge of large language models and the adaptable nature of cognitive architectures. More concretely, we formally define a cognitive architecture, show how we can bootstrap its rules with a large language model and minimal human input, collect a set of human preferences in the real world, and show how the architecture we proposed can adapt to those preferences in one shot. We hope that this work will inspire more future work in rule-centric agent policies in the future.

66 pages

Advisor
Reid Simmons

Return to: SCS Technical Report Collection
School of Computer Science

This page maintained by reports@cs.cmu.edu