Research map / Learning theory and algorithms / Prediction and decision losses
Gradient Temporal-Difference Learning with Regularized Corrections
Sina Ghiassian, Andrew Patterson, Shivam Garg, Dhawal Gupta, Adam White, Martha White
Read the paperCheck your idea against 137,217 papers
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 |
|---|---|---|---|
| Add ℓ2 regularization to the secondary-weight update (TDRC instead of TDC) (ablation) | TDRC · Boyan's Chain | RMSPBE area: 0.639 ± 0.001 -> 0.222 ± 0.002 | Improved |
| Add ℓ2 regularization to the secondary-weight update (TDRC instead of TDC) (ablation) | TDRC · Baird's Counterexample | RMSPBE area: 0.196 ± 0.007 -> 0.242 ± 0.006 | Mixed |
| Add ℓ2 regularization to the secondary-weight update (TDRC instead of TDC) (ablation) | TDRC · Random Walk (Inverted) | RMSPBE area: 0.053 ± 0.001 -> 0.047 ± 0.001 | Improved |
| Add ℓ2 regularization to the secondary-weight update (TDRC instead of TDC) (ablation) | TDRC · Random Walk (Tabular) | RMSPBE area: 0.063 ± 0.001 -> 0.049 ± 0.001 | Mixed |
| Increase regularization from zero (TDC) to a positive β (ablation) | TDRC · Random Walk, Boyan's Chain, and Baird's Counterexample | RMSPBE | Improved |
| Use β = 1.0 rather than the better-performing larger β values identified by tuning (ablation) | TDRC · Linear prediction problems | RMSPBE | Hurt |
| Reduce the secondary-weights stepsize multiplier to η = 0.5 (ablation) | QC · Mountain Car (neural network) | steps to goal | Improved |
| Use QC instead of Q-learning | QC · MinAtar Breakout and Space Invaders | average return | Improved |
| Use QRC with β = 1 instead of QC (ablation) | QRC · MinAtar Breakout and Space Invaders | average return | Hurt |
| Use QRC instead of Q-learning | QRC · Mountain Car with 10 replay steps | steps to goal | Improved |
| Add target networks (ablation) | Q-learning · Non-linear control environments | learning performance | Hurt |
| Use momentum with Adam instead of Adagrad (ablation) | TDRC and comparison methods · Linear prediction problems | RMSPBE | No clear effect |
Closest papers by meaning
- Accelerated Gradient Temporal Difference Learning (AAAI 2017 · similarity 0.89)
- Towards Parameter-Free Temporal Difference Learning (ICML 2026 · similarity 0.88)
- Almost Sure Convergence of Linear Temporal Difference Learning with Arbitrary Features (NeurIPS 2026 · similarity 0.87)
- Discerning Temporal Difference Learning (AAAI 2024 · similarity 0.86)
- On Double Descent in Reinforcement Learning with LSTD and Random Features (ICLR 2024 · similarity 0.86)
- An Improved Finite-time Analysis of Temporal Difference Learning with Deep Neural Networks (ICML 2024 · similarity 0.85)
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.
