Research map / Reinforcement learning / Policy gradient and actor-critic
Beyond the Bellman Recursion: A Pontryagin-Guided Framework for Non-Exponential Discounting
Hojin Ko, Jeonggyu Huh
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 |
|---|---|---|---|
| use PG-DPO instead of DPO (add Stage 2 Adjoint-MC projection) (ablation) | PG-DPO · Case 1 survival discounting (beta0=0.2) | Global L1 error: 3.82e-2 -> 1.45e-2 | Improved |
| add Stage 2 Adjoint-MC projection to Stage 1 DPO (ablation) | DPO · Case 2 hyperbolic discounting (Consumption) | L1 error (MAE): 2.32e-01 -> 3.47e-03 | Improved |
| add Stage 2 Adjoint-MC projection to Stage 1 DPO (ablation) | DPO · Case 2 hyperbolic discounting (Investment) | L1 error (MAE): 3.12e-01 -> 6.36e-08 | Improved |
| add Stage 2 Adjoint-MC projection to Stage 1 DPO (ablation) | DPO · Case 3 time-varying hyperbolic (Sinusoidal k(t)) | Global L1 error: 1.31e-01 -> 7.39e-03 | Improved |
| use PG-DPO instead of PPO actor-critic | PG-DPO · Case 1 survival discounting (beta0=0.5) | Global L1 error: 2.18e-1 -> 2.97e-2 | Improved |
| use PG-DPO instead of PINN | PG-DPO · Case 1 survival discounting (beta0=1.0) | Global L1 error: 1.65e0 -> 1.80e-2 | Improved |
| use PG-DPO instead of BSVIE | PG-DPO · Case 2 hyperbolic discounting (Investment) | L-inf error (max): 2.44e-01 -> 1.92e-06 | Improved |
| use PG-DPO instead of Deep BSVIE | PG-DPO · Case 3 time-varying hyperbolic (Linear k(t)) | Global L1 error: 1.29e0 -> 6.35e-03 | Improved |
| use PG-DPO instead of PPO | PG-DPO · Case 3 time-varying hyperbolic (Exponential k(t)) | Global L1 error: 2.42e-01 -> 6.70e-03 | Improved |
| PINN vs Deep BSDE as global equation-driven baseline at small beta0 | Deep BSDE · Case 1 survival discounting (beta0=0.2) | Global L1 error: 1.11e0 -> 1.95e-1 | Improved |
Closest papers by meaning
- Exploratory Preference Optimization: Harnessing Implicit Q*-Approximation for Sample-Efficient RLHF (ICLR 2025 · similarity 0.87)
- Natural Policy Gradient as Doubly Smoothed Policy Iteration: A Bellman-Operator Framework (NeurIPS 2026 · similarity 0.86)
- Zeroth-Order Policy Gradient for Reinforcement Learning from Human Feedback without Reward Inference (ICLR 2025 · similarity 0.86)
- A Continuous-time Tractable Model for Present-biased Agents (AAAI 2025 · similarity 0.86)
- Follow the Winners: Conservative Policy Improvement with the Cross-Entropy Method for Critic-Free RFT (NeurIPS 2026 · similarity 0.86)
- Reward Model Learning vs. Direct Policy Optimization: A Comparative Analysis of Learning from Human Preferences (ICML 2024 · 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.
