Research map / Learning theory and algorithms / Submodular and game algorithms
ESCHER: Eschewing Importance Sampling in Games by Computing a History Value Function to Estimate Regret
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Estimate regret directly with the history value function and omit importance sampling | ESCHER · Phantom Tic-Tac-Toe, Dark Hex 4, Dark Hex 5 | regret estimator variance: ESCHER: (2.6 ± 0.1) × 10^{-1}, (1.8 ± 0.1) × 10^{-1}, (1.3 ± 0.1) × 10^{-1}; DREAM: (4.6 ± 1.0) × 10^{7}, (2.8 ± 2.0) × | Helped |
| Add a bootstrapped history-value baseline (ablation) | ESCHER · Dark Hex 4 | regret estimator variance: (1.8 ± 0.1) × 10^{-1} -> (1.3 ± 0.9) × 10^{2} | Hurt |
| Add reach weighting (ablation) | ESCHER · Dark Hex 5 | regret estimator variance: (1.3 ± 0.1) × 10^{-1} -> (2.0 ± 0.6) × 10^{5} | Hurt |
| Add a bootstrapped history-value baseline (ablation) | ESCHER · Phantom Tic-Tac-Toe and Dark Hex 5 | regret estimator variance: Phantom Tic-Tac-Toe: (2.6 ± 0.1) × 10^{-1} -> (4.1 ± 0.7) × 10^{1}; Dark Hex 5: (1.3 ± 0.1) × 10^{-1} -> (3.3 ± 1.6) × 1 | Hurt |
| Add reach weighting (ablation) | ESCHER · Phantom Tic-Tac-Toe, Dark Hex 4, Leduc, Battleship, and Liar's Dice | regret estimator variance: Phantom Tic-Tac-Toe: (2.6 ± 0.1) × 10^{-1} -> (1.4 ± 0.4) × 10^{7}; Dark Hex 4: (1.8 ± 0.1) × 10^{-1} -> (3.1 ± 1.7) × 1 | Hurt |
| Use ESCHER rather than DREAM | ESCHER · Dark Chess | head-to-head win rate: over $90\%$ | Helped |
| Use ESCHER rather than DREAM or NFSP | ESCHER · Phantom Tic-Tac-Toe | head-to-head performance | No clear effect |
| Use ESCHER rather than DREAM or NFSP | ESCHER · Dark Hex 5 | head-to-head performance | Helped |
Closest papers by meaning
- Double Neural Counterfactual Regret Minimization (ICLR 2020 · similarity 0.90)
- Learning Not to Regret (AAAI 2024 · similarity 0.90)
- Learning Deviation Payoffs in Simulation-Based Games (AAAI 2019 · similarity 0.89)
- Deep (Predictive) Discounted Counterfactual Regret Minimization (AAAI 2026 · similarity 0.88)
- Optimism Without Regularization: Constant Regret in Zero-Sum Games (NeurIPS 2025 · similarity 0.87)
- Preference-CFR: Beyond Nash Equilibrium for Better Game Strategies (ICML 2025 · similarity 0.87)
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.