Summary of Decision Mamba: a Multi-grained State Space Model with Self-evolution Regularization For Offline Rl, by Qi Lv et al.
Decision Mamba: A Multi-Grained State Space Model with Self-Evolution Regularization for Offline RL
by Qi Lv, Xiang Deng, Gongwei Chen, Michael Yu Wang, Liqiang Nie
First submitted to arxiv on: 8 Jun 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 The abstract proposes Decision Mamba (DM), a novel multi-grained state space model designed for offline reinforcement learning (RL) tasks. The transformer-based conditional sequence modeling struggles to handle out-of-distribution states and actions, which existing work attempts to address using data augmentation or value-based RL algorithms. However, these approaches still face challenges such as insufficient utilization of historical temporal information, overlooking local relationships among return-to-gos, states, and actions, and overfitting suboptimal trajectories with noisy labels. DM addresses these issues by proposing a self-evolving policy learning strategy that incorporates the mamba architecture to model the historical hidden state and fine-grained SSM module to capture the RTG-state-action relationships. Additionally, progressive regularization is used to mitigate overfitting on noisy trajectories. The proposed approach is shown to outperform other baselines in extensive experiments. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Decision Mamba (DM) is a new way to solve offline reinforcement learning problems. Right now, we have models that are good at certain tasks, but they struggle when things don’t go exactly as planned. DM tries to fix this by looking at the past and learning from its mistakes. It also focuses on small details within each step, like what happens when a robot returns to a previous spot. This approach is different from current methods because it uses a combination of old and new information to make better decisions. The results show that DM performs much better than other approaches in various tasks. |
Keywords
» Artificial intelligence » Data augmentation » Overfitting » Regularization » Reinforcement learning » Transformer
