On-Policy Vs. Off-Policy Reinforcement Learning in ConnectX: Seat-Stratified Performance and the Role of Action Masking
Bibliographic record
Abstract
This study evaluates on-policy and off-policy reinforcement learning for ConnectX under a unified, seat-stratified evaluation protocol, with particular attention to invalid-action masking. Policies are trained against a fixed opponent schedule and evaluated offline with frozen parameters. The test set distinguishes Seen opponents from Unseen, deeper search (noisy-minimax-3 as the primary adversary; clean minimax-3 as a stress test). Uncertainty is summarized with 95% Wilson intervals, and all outcomes are reported by seat to avoid pooling bias. Across Seen opponents, masking consistently accelerates policy-gradient learning and reduces variability: MaskablePPO reaches target win-rate levels with fewer samples and tighter intervals than PPO, while value-based baselines reach an early asymptote. Against the Unseen noisy-minimax-3 in the first seat, PPO exceeds MaskablePPO (71.9% vs. 37.5%; intervals non-overlapping), whereas in the second seat masking yields large margins where legal actions are scarce (e.g., 87.6% vs. 17.3% against noisy-minimax-2; 48.3% vs. 8.3% against blocker; intervals non-overlapping). Against clean minimax-3, win rates approach zero within CI in the second seat for all methods, indicating a persistent gap to deeper search. These results imply two practical guidelines for legality-constrained, turn-based games: (i) legality-aware sampling is a substantive stability lever for policy-gradient methods rather than a peripheral implementation detail; and (ii) seat-stratified reporting is required to disentangle structural initiative from algorithmic effects. Overall, masking improves in-distribution efficiency and second-seat robustness but does not by itself close the generalization gap to stronger, unseen search.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.005 | 0.015 |
| Meta-epidemiology (narrow) | 0.002 | 0.000 |
| Meta-epidemiology (broad) | 0.001 | 0.000 |
| Bibliometrics | 0.000 | 0.000 |
| Science and technology studies | 0.000 | 0.002 |
| Scholarly communication | 0.001 | 0.002 |
| Open science | 0.002 | 0.002 |
| Research integrity | 0.002 | 0.003 |
| Insufficient payload (model declined to judge) | 0.003 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".