![[2602.19373] Stable Deep Reinforcement Learning via Isotropic Gaussian Representations](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.19373] Stable Deep Reinforcement Learning via Isotropic Gaussian Representations
Summary
This paper presents a method for enhancing stability in deep reinforcement learning by utilizing isotropic Gaussian representations, addressing issues of non-stationarity in training dynamics.
Why It Matters
The research is significant as it tackles the prevalent challenge of unstable training in deep reinforcement learning systems, which can hinder performance in dynamic environments. By proposing a novel regularization technique, the authors provide a practical solution that could enhance the adaptability and reliability of AI agents across various applications.
Key Takeaways
- Isotropic Gaussian embeddings improve stability in reinforcement learning.
- The proposed method reduces representation collapse and neuron dormancy.
- Empirical results show enhanced performance in non-stationary environments.
- The approach is computationally inexpensive and easy to implement.
- Stable tracking of time-varying targets is achieved with linear readouts.
Computer Science > Machine Learning arXiv:2602.19373 (cs) [Submitted on 22 Feb 2026] Title:Stable Deep Reinforcement Learning via Isotropic Gaussian Representations Authors:Ali Saheb, Johan Obando-Ceron, Aaron Courville, Pouya Bashivan, Pablo Samuel Castro View a PDF of the paper titled Stable Deep Reinforcement Learning via Isotropic Gaussian Representations, by Ali Saheb and 4 other authors View PDF HTML (experimental) Abstract:Deep reinforcement learning systems often suffer from unstable training dynamics due to non-stationarity, where learning objectives and data distributions evolve over time. We show that under non-stationary targets, isotropic Gaussian embeddings are provably advantageous. In particular, they induce stable tracking of time-varying targets for linear readouts, achieve maximal entropy under a fixed variance budget, and encourage a balanced use of all representational dimensions--all of which enable agents to be more adaptive and stable. Building on this insight, we propose the use of Sketched Isotropic Gaussian Regularization for shaping representations toward an isotropic Gaussian distribution during training. We demonstrate empirically, over a variety of domains, that this simple and computationally inexpensive method improves performance under non-stationarity while reducing representation collapse, neuron dormancy, and training instability. Subjects: Machine Learning (cs.LG); Artificial Intelligence (cs.AI) Cite as: arXiv:2602.19373 [cs.LG] (or...