Solving the Rubik’s Cube in a Human-like Manner with Assisted Reinforcement Learning (ARL)

AAAI 2025 Workshop NeurMAD Submission24 Authors

Keywords: ai, ml, rl, inverse rl, deep learning

TL;DR: Our work explores the field of human-AI systems and an example of how to make AI more interpretable and user-friendly with the Rubik's cube/reinforcement learning.

Abstract:

Human-AI collaboration is most key in situations in which AI must approach problems in a human-like manner. In this work, we present a novel approach to Rubik’s cube solving that utilizes human-like solving techniques. We demonstrate assisted reinforcement learning (ARL), in which RL trains to solve the cube in separate steps (CFOP), thereby emulating human behavior.
Secondly, we applied inverse reinforcement learning (IRL) to align AI behavior with human problem-solving. We create a dataset of over 10,000 human Rubik’s cube solves and train to achieve a reward function that accurately reflects the goals and preferences of human solvers. As a result, the system is able to generalize across different cube states while maintaining interpretability.
Our research demonstrates the potential of combining ARL and IRL to close the gap between human and AI behavior. We successfully highlight the interdisciplinary nature of training AI to solve a trivial task while imitating complex human behavior.

Submission Number: 24