Research map / Optimization / Optimizers
Resetting the Optimizer in Deep RL: An Empirical Study
Kavosh Asadi, Rasool Fakoor, Shoham Sabach
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Reset Adam's moment estimates at each target-network update | Rainbow with Adam · Atari, 12 games | human-normalized median score | Improved |
| Reset Adam's moments stochastically with probability 1/K after updates (ablation) | Rainbow with Adam · Atari, 12 games | human-normalized performance | Improved |
| Reset at each target update instead of resetting stochastically (ablation) | Rainbow with Adam · Atari, 12 games | human-normalized performance | Improved |
| Use K=1, so moments reset after every step (ablation) | Rainbow with Adam · Atari, 12 games | human-normalized median score | Hurt |
| Use smaller K values rather than the default K=8000 while resetting (ablation) | Rainbow with Adam · Atari, 12 games | human-normalized performance | Improved |
| Reset RMSProp at each target-network update | Rainbow with RMSProp · Atari, 12 games | human-normalized median score | Improved |
| Reset Rectified Adam at each target-network update | Rainbow with Rectified Adam · Atari, 12 games | human-normalized median score | Improved |
| Reset Adam at each target-network update | DQN with Adam · Atari, 12 games | human-normalized median score | Improved |
| Reset Adam at each target-network update | RainbowPro with Adam · Atari, 12 games | human-normalized median score | Improved |
| Reset only Adam's first-order moment (ablation) | Rainbow with Adam · Atari, 12 games | human-normalized performance | No clear effect |
| Reset only Adam's second-order moment (ablation) | Rainbow with Adam · Atari, 12 games | human-normalized performance | Improved |
| Reset both Adam moments rather than only one (ablation) | Rainbow with Adam · Atari, 12 games | human-normalized performance | Improved |
Closest papers by meaning
- Implicit Under-Parameterization Inhibits Data-Efficient Deep Reinforcement Learning (ICLR 2021 · similarity 0.89)
- Adam on Local Time: Addressing Nonstationarity in RL with Relative Adam Timesteps (NeurIPS 2024 · similarity 0.87)
- The Effective Horizon Explains Deep RL Performance in Stochastic Environments (ICLR 2024 · similarity 0.87)
- Towards General-Purpose Model-Free Reinforcement Learning (ICLR 2025 · similarity 0.87)
- Stop Regressing: Training Value Functions via Classification for Scalable Deep RL (ICML 2024 · similarity 0.87)
- MAD-TD: Model-Augmented Data stabilizes High Update Ratio RL (ICLR 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
