Programmatic policies for interpretable reinforcement learning using pre-trained models
Decision Trees (DTs) are widely used in machine learning due to their critical interpretability. However, training DTs in a Reinforcement Learning (RL) setting is challenging. In the project, we present a framework to improve the interpretability of reinforcement learning (RL) by generating prog...
Saved in:
| 主要作者: | |
|---|---|
| 其他作者: | |
| 格式: | Final Year Project |
| 語言: | English |
| 出版: |
Nanyang Technological University
2024
|
| 主題: | |
| 在線閱讀: | https://hdl.handle.net/10356/181169 |
| 標簽: |
添加標簽
沒有標簽, 成為第一個標記此記錄!
|
| 機構: | Nanyang Technological University |
| 語言: | English |
| 總結: | Decision Trees (DTs) are widely used in machine learning due to their critical
interpretability. However, training DTs in a Reinforcement Learning (RL) setting
is challenging.
In the project, we present a framework to improve the interpretability of
reinforcement learning (RL) by generating programmatic policies through large
language models (LLMs). Initially looking at decision trees within latent space
representation, we proceed to develop and apply a “code reflection” framework
in the Karel environment, which provides a controlled setting for evaluating task
performance. The framework leverages prompt engineering and code reflection
to optimize program synthesis, producing interpretable policies.
The report discusses the findings and concludes with recommendations for future
work at the end. |
|---|