Summary of Beyond Simple Sum Of Delayed Rewards: Non-markovian Reward Modeling For Reinforcement Learning, by Yuting Tang et al.
Beyond Simple Sum of Delayed Rewards: Non-Markovian Reward Modeling for Reinforcement Learning
by Yuting Tang, Xin-Qiang Cai, Jing-Cheng Pang, Qiyu Wu, Yao-Xiang Ding, Masashi Sugiyama
First submitted to arxiv on: 26 Oct 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: None
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary Reinforcement Learning (RL) agents learn from reward signals to acquire skills. Designing high-quality instance-level rewards is labor-intensive. An alternative, RL with delayed rewards, focuses on learning from periodic rewards obtained from human evaluators assessing sequences of behaviors. However, traditional methods assume underlying Markovian rewards and a sum of instance-level rewards, which don’t align well with real-world scenarios. This paper introduces RLCoDe, generalizing traditional RL from delayed rewards by eliminating the strong assumption. We present a framework modeling composite delayed rewards using weighted sums of non-Markovian components to capture individual step contributions. We propose CoDeTr, incorporating in-sequence attention to effectively model these contributions. Experiments on challenging locomotion tasks show CoDeTr outperforms baseline methods across evaluated metrics, accurately identifying significant time steps and predicting rewards reflecting environment feedback. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper is about teaching machines to learn from rewards that come later. Right now, it’s hard to design good rewards for machines to learn. A new way of learning involves getting rewards periodically, but this can be tricky too. The problem is that some methods assume the rewards are simple and easy to understand, which isn’t true in real life. This paper introduces a new way of looking at delayed rewards, using a combination of different things that happen over time. It also proposes a special machine learning model called CoDeTr that can learn from these complex rewards. The results show that this model works better than others and can even identify the most important parts of what’s happening. |
Keywords
* Artificial intelligence * Attention * Machine learning * Reinforcement learning
