# Efficient Learning in Chinese Checkers: Comparing Parameter Sharing in Multi-Agent Reinforcement Learning

**Authors:** Noah Adhikari, Allen Gu  
**arXiv:** [2405.18733](https://arxiv.org/abs/2405.18733)  
**Date:** 2024-05-29  
**Summary page:** https://www.summarizepaper.com/en/paper/2405.18733v1/

## Key points

- Allen Gu and Noah Adhikari from the University of California, Berkeley, explore efficiency of multi-agent reinforcement learning (MARL) with full parameter sharing in Chinese Checkers
- Their approach outperforms independent and partially shared architectures in competitive perfect-information homogenous game
- Researchers develop a novel MARL environment tailored for variable-size, six-player Chinese Checkers within PettingZoo
- Environment adheres to traditional rules including chaining jumps, making it faithful to the true essence of Chinese Checkers
- Challenges in Chinese Checkers include large branching factor and potentially infinite horizons
- Gu and Adhikari introduce concept of branching actions (submoves) to reduce dimensionality of action space
- Observation space inspired by AlphaGo using multiple binary game boards stacked in 3D array
- Research team provides access to PettingZoo environment, training procedures, evaluation logic, and analysis scripts on Github for further exploration and replication

## Summary

In their study, Allen Gu and Noah Adhikari from the University of California, Berkeley, explore the efficiency of multi-agent reinforcement learning (MARL) with full parameter sharing in the context of Chinese Checkers. They demonstrate that this approach outperforms independent and partially shared architectures in the competitive perfect-information homogenous game of Chinese Checkers. To conduct their experiments, the researchers develop a novel MARL environment specifically tailored for variable-size, six-player Chinese Checkers. This custom environment, built within PettingZoo, adheres to all traditional rules of the game including chaining jumps. This makes it the first implementation that remains faithful to the true essence of Chinese Checkers. Chinese Checkers poses significant challenges for learning algorithms due to its large branching factor and potentially infinite horizons. To address this complexity, Gu and Adhikari introduce the concept of branching actions (submoves) borrowed from other reinforcement learning domains with complex action spaces. By allowing submoves that do not immediately end a player's turn, they effectively reduce the dimensionality of the action space. Furthermore, their observation space is inspired by AlphaGo, utilizing multiple binary game boards stacked in a 3D array to encode crucial information. The research team provides access to their PettingZoo environment, training procedures, evaluation logic, and analysis scripts on Github for further exploration and replication by interested parties. Through their comprehensive study and innovative approaches to tackling challenges in Chinese Checkers through MARL techniques with full parameter sharing,<Organization>, contribute valuable insights to advancing efficient learning strategies in complex board games like Chinese Checkers.