# Evolutionary learning of interpretable decision trees

**Authors:** Leonardo Lucio Custode, Giovanni Iacca  
**arXiv:** [2012.07723](https://arxiv.org/abs/2012.07723)  
**Date:** 2020-12-14  
**Summary page:** https://www.summarizepaper.com/en/paper/2012.07723v1/

## Key points

- Reinforcement learning techniques have achieved human-level performance in various tasks over the past decade.
- Interpretability has become increasingly important as the field advances.
- Understanding how AI systems operate and the rationale behind their decisions is crucial for safety assessment and problem-solving.
- Existing techniques for interpretable reinforcement learning may be limited by suboptimal solutions due to reliance on greedy algorithms or underutilization of rewards.
- A novel approach using decision trees combines evolutionary algorithms with Q-learning for enhanced performance and interpretability.
- The two-level optimization strategy aims to improve state space decomposition and action association, leading to competitive results in benchmark tests.
- Experimentation showed that the new method is on par with current state-of-the-art techniques in terms of both performance and interpretability.
- Ablation study confirmed significant performance improvement in complex environments compared to traditional single-layer optimization methods.

## Summary

In the past decade, reinforcement learning techniques have made remarkable progress in achieving human-level performance across various tasks. However, as the field has advanced, the importance of interpretability has become increasingly evident. The ability to comprehend how a system operates and understand the rationale behind its decisions is crucial not only for assessing the safety of AI systems but also for extracting valuable insights into previously unknown problems. While existing techniques optimize decision trees for reinforcement learning, they often rely on greedy algorithms or fail to fully leverage the rewards provided by the environment. This limitation can lead to these methods becoming trapped in suboptimal solutions. To address this challenge, a novel approach to interpretable reinforcement learning using decision trees has been proposed. This innovative approach adopts a two-level optimization strategy that combines the strengths of evolutionary algorithms with those of Q-learning. By breaking down the problem into two distinct sub-problems – identifying a meaningful state space decomposition and associating actions with each state – this method aims to enhance both performance and interpretability. The effectiveness of this approach was evaluated through experimentation on three well-known reinforcement learning benchmarks. The results demonstrate that this new method is competitive with current state-of-the-art techniques in terms of both performance and interpretability. Additionally, an ablation study confirmed that utilizing the two-level optimization approach provides a significant performance boost in complex environments compared to traditional single-layer optimization techniques. Overall, this research contributes valuable insights into advancing interpretable reinforcement learning methods, offering a promising avenue for developing more transparent and effective AI systems in the future.