Node-lock costs in a three-player postflop game: direct changes and re-adaptation
A node lock fixes one player's strategy at a single decision and asks what the restriction costs. This note separates the direct cost D of a lock, with every other policy held at the reference, from the total cost C after all three seats re-solve, in a 40 bb three-player postflop spot solved with discounted counterfactual regret minimisation. In an archived panel of 53 locks on J♥9♦8♦, built by moving first the hands that lose least from the move, the residual A = C − D accounts for 89% and 71% of summed cost for ±10 and ±20 percentage-point moves, but one lock carries 45% of the ±10 pp total and the median lock-level shares are 81% and 73%. The opponents' unilateral best-response gains exceed C for every archived lock, about tenfold at the median. Version 1.1 adds runs whose analysis was fixed before they were made. When the same aggregate move is spread over all hands or made with hands in random order, summed D rises about fourfold. In five re-solves with these constructions, C changes in both directions; for the dominant lock it falls from 0.395 bb to 0.0007 bb, too small to distinguish from zero in these solves. A small direct cost is therefore a property of the lock construction, not of node locking in general; in the five re-solves, C and its split into D and A also change with the construction. With opponents held at the reference, the locked player's own re-optimisation avoids 77–87% of summed D on J♥9♦8♦, mostly by reaching the locked node less often. The results describe these locks in one modelled spot. The record contains the note (PDF and Markdown), the data bundle (tables, lock lists, the version 1.1 runs and checks, analysis scripts, push/fold dataset and run commands; CC BY 4.0) and the solver source code mtt-cfr 0.6.0 (Rust; MIT). The reference policies (about 2.1 GB each) are not included and will be provided with a later release.
Authors
- Nari Han (ORCID: https://orcid.org/0009-0008-7228-8740)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-14
- DOI
- https://doi.org/10.5281/zenodo.24212414
- Primary Topic
- Artificial Intelligence in Games
- Type
- article
- Field-Weighted Citation Impact
- 0.00