acceptodds
Under review as a conference paper at ICLR 2027

ONE SEARCH, TWO VALUES: TEST-TIME SEARCH EFFICIENCY IN THREE-PLAYER GENERAL-SUM GAMES

Abstract

With three or more players, a search should back up the value aligned with the players’ objective. Objectives such as win probability or rank credit make that aligned value coarse: a player’s actions are mostly exactly tied or far apart, so a few-simulation search over the aligned value is often left undecided. The cardinal score is dense and decisive but misaligned, since the score-optimal action can be strictly worse in win probability. We carry both values through one search. A tree search guided by the objective-aligned value head also backs up the cardinal head’s vector through the same nodes, so both action-value vectors reach the root with no additional search simulation. A gated readout then plays the aligned decision where the aligned margin clears a calibrated gate threshold and the cardinal ordering where it does not. The readout leaves the tree unchanged, so its effect is confined to the states where it fires, and its expected gain is the firing probability times the accuracy difference of the two readouts on those states, which we measure. On an exactly solved eight-token ranking game and on a reduced three-player Chinese Checkers, the gated readout at 64 simulations per move beats the aligned readout at 64 and 128 simulations and the cardinal search at both budgets. The aligned readout needs about seven times the simulation budget to match it on the ranking game under the rank objective, and on Chinese Checkers it does not match it within 1024 simulations.

Primary area:
reinforcement learning
Keywords:
Monte Carlo Tree Search, Test-Time Planning, Multi-Player Games, Value-Guided Search, Reinforcement Learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_emP9OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper focuses on three-player perfect-information games where the game objective mismatches the cardinal score. When the optimization objectives are win probability or rank credit, the aligned value can be coarse; searching with few simulations over the aligned value cannot reliably decide root actions. Cardinal values like the raw score can provide more information, but can mismatch the aligned value. To handle this, this paper proposes to carry both aligned values and cardinal values through one search. The value outputs share the same network body, so one forward pass can directly output both values. In a single PUCT search guided by the aligned value, the method independently performs max^n backup for both aligned and cardinal value vectors. Then, the method decides whether to use aligned readout or cardinal readout based on whether the aligned margin, i.e., the gap between the top two aligned action values, clears a calibrated threshold. The experiments use exactly solved Tally and reduced Chinese Checkers with an exactly solvable endgame band, demonstrating that under 64 simulations, gated readout performs better than aligned readout and cardinal search.

Critical strengths

The empirical analysis is quite comprehensive, and the evaluation is rigorous. This includes the tournament credit, the regret calculated by the exact max^n solver for Tally, and also the analysis of the decision quality in Chinese checkers solvable endgames. Figure 2 presents the regret of the two readouts, giving the gated readout a direct empirical motivation.

Critical weaknesses

The novelty of the dual-backup search and the margin gate remains unclear. Implicit minimax backups (Lanctot et al., 2014) already propose to maintain and backup two evaluation signals within a single MCTS tree, but this work lacks a discussion or comparison with it. KataGo also studies using win signal and score together; this work compares only with a KataGo-style fixed linear blend, which may not reflect KataGo's actual score utility.

The comparisons cannot fully isolate the source of performance gain. The effectiveness of tree guidance is supported by showing that always-cardinal readout performs better than cardinal search. However, the two methods differ in both the value head guiding the tree and the root action selection (value argmax vs. most-visited), and Appendix F suggests the selection rule alone can affect regret. The main test-time baselines also use PUCT most-visited readout, without comparing with Gumbel MuZero's sequential halving and corresponding root action selection under the same simulation budget.

Questions for authors
  1. Please clarify the advantages over prior works like the two mentioned above. Are their approaches not working in the target scenarios? Such as implicit minimax backups or KataGo’s actual score utility beyond the fixed linear blend.
  2. Is it possible to do more test-time comparisons to isolate the performance gain? Such as adding one to use cardinal search with argmax cardinal value readout, compared with the current always-cardinal readout.
  3. Can Gumbel MuZero's sequential halving and root action selection serve as test-time comparison baselines? This can help confirm whether the proposed method performs better than the existing few-simulation search.
  4. In the appendix, the experiments show that the gain declines with training. Does this reflect that cardinal values provide a target that is easier to learn than aligned values? So, on the other hand, is it simply an intermediate auxiliary before the aligned values learn?
8 comments

Dear Reviewer emP9,

Thank you for the careful review and the suggested controls. We ran each of them and report the results below, including the settings in which a control is stronger than our method. All comparisons concern test-time search in Tally and reduced Chinese Checkers (CC), with fixed networks and opponents trained using replacement backup.

Comment 1. Implicit minimax backups and KataGo

W1 / Q1. Novelty relative to implicit minimax backups (IMB) and KataGo's score utility.

The difference lies in where the second value is used. IMB and KataGo's score utility change the values that guide tree selection, whereas our method leaves the tree unchanged and uses the cardinal value only for final action selection, when the aligned margin is below the threshold.

MethodValue guiding tree selectionFinal action
IMB PUCTAligned mean return mixed with a heuristic value backed up by replacement (weight α\alpha)Most visited
KataGo-style utilityAligned value plus a bounded function of the predicted score leadMost visited
Gated (ours)Aligned value only; the cardinal vector is backed up independently and does not affect selectionHighest backed-up cardinal value when the aligned margin is below the threshold; most visited otherwise

The rows for IMB and the KataGo-style utility describe the controls we implemented (Appendices K and L).

The final-action rule follows from a measurement (§3, Figure 2). On solved Tally positions, the aligned readout has its largest regret at small, nonzero aligned margins, where the cardinal readout is more often better. The gate therefore uses the cardinal readout in this range and retains the aligned readout elsewhere.

To test whether the two prior approaches work in our setting, we added IMB PUCT with either head as its heuristic (weight 0.5) and a three-player adaptation of KataGo's nonlinear score utility. The latter uses a root-centered arctangent of the predicted lead over the strongest opponent, with weight 0.2 on our [0,1][0,1] payoff scale. Appendix L specifies the adaptation, which evaluates the utility at the predicted mean lead.

Gated denotes our original replacement-backup search with the gate. Unless stated, test networks are trained on 300 self-play games. Entries compare tournament credit against the same two copies of a fixed, 300-game win-trained opponent; intervals are paired, pointwise 95% bootstrap intervals over tables or openings. Bold marks intervals excluding zero.

Gated at 64 simulations minusTally rankTally winCC rankCC win
IMB, aligned heuristic, 64+.006 [-.000,.012]+.000 [-.007,.007]+.245 [.218,.271]+.342 [.306,.378]
IMB, cardinal heuristic, 64+.008 [.002,.014]-.003 [-.010,.004]+.155 [.136,.173]+.241 [.202,.280]
KataGo-style utility, 64+.002 [-.004,.008]-.010 [-.017,-.003]+.157 [.138,.176]+.215 [.170,.260]

Both approaches work in our setting: cardinal-heuristic IMB improves mean-backup PUCT on CC, and the KataGo-style utility improves it on both games. On CC, Gated at 64 simulations further exceeds all three controls at both 64 and 128 simulations. On Tally rank, Gated exceeds cardinal-heuristic IMB by +.008 [.002,.014] at 64 simulations, while the KataGo-style utility is higher on Tally win. At 128 simulations, all three controls exceed Gated at 64 on Tally win.

Manuscript locations: §3, Figure 2, §5.3, Tables 2 and 13, §6 and Appendices K–L.

Comment 2. Separating tree guidance from final action selection

W2 / Q2. The always-cardinal readout and cardinal search differ in both guidance and final action selection.

The original comparison changed both factors. We added the requested cardinal-argmax control: cardinal-guided search with its tree unchanged and its final action chosen by cardinal value instead of visit count. This separates guidance with final action selection fixed, and final action selection with guidance fixed. At 64 simulations:

ComparisonTally rankTally winCC rankCC win
Guidance, final action selection fixed: always-cardinal minus cardinal-argmax+.002 [-.002,.007]+.006 [.002,.010]-.001 [-.018,.016]-.022 [-.055,.011]
Final action selection, guidance fixed: cardinal-argmax minus cardinal search+.007 [.001,.012]+.008 [.001,.014]+.106 [.087,.124]+.206 [.169,.242]

Final action selection accounts for most of the original difference; aligned guidance adds a resolved gain on Tally win. Selective switching contributes separately on Tally: Gated exceeds the always-cardinal readout on the same tree by +.011 (rank) and +.007 (win). On CC, the gate fires on about 99% of moves, and the difference from cardinal-argmax search is unresolved (−.002, −.026).

Manuscript locations: §5.4, Table 1 and Appendix J.

Comment 3. Gumbel search and the gate on a mean-backup tree

W2 / Q3. Comparison with Gumbel sequential halving and its root selection at the same budget.

As requested, we compared the submitted Gated method with Gumbel searches using mean and replacement backups. These controls follow the official mctx selection equations, including root sequential halving, mixed-value completion and the survivor-restricted recommendation. GM denotes mean-backup Gumbel with no root noise; noise scale 1 is tested in Appendix K.

Gated at 64 simulations minusTally rankTally winCC rankCC win
GM, 64-.013 [-.020,-.006]-.026 [-.033,-.019]+.078 [.055,.100]+.207 [.168,.246]
GM, 128-.018 [-.026,-.010]-.026 [-.033,-.019]+.032 [.009,.053]+.145 [.110,.181]
Gumbel, replacement backup, 64+.005 [-.003,.013]+.008 [.0004,.015]+.147 [.124,.169]+.281 [.247,.316]

With the 300-game networks, Gated at 64 simulations exceeds GM at both 64 and 128 simulations on both CC objectives; GM is stronger on Tally.

In an additional experiment, we applied the same final-action rule to GM itself. GM+gate keeps GM's mean aligned backup and tree, independently backs up the cardinal vector by replacement, and uses the same threshold 2Δ^A2\hat\Delta_A, retaining GM's own recommendation when the gate does not fire. All configurations below use 64 simulations.

ComparisonTally rankTally winCC rankCC win
GM+gate − GM-.003 [-.009,.002]-.006 [-.011,-.002]+.097 [.077,.116]+.269 [.227,.307]
GM+gate − Gated+.010 [.003,.016]+.019 [.013,.027]+.019 [.004,.034]+.061 [.034,.090]

GM+gate exceeds the original Gated configuration in all four settings. On Tally, its point estimates are below GM. On CC, its improvement over GM shows that the proposed final-action construction can also improve mean-backup Gumbel search.

Manuscript locations: §5.3–§5.4, Tables 2, 13 and 23, and Appendix L, Tables 30–31.

Reviewer_c8aVOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
2(+1)
Read review
Paper summary

The paper claims to improve some kind of search algorithm (see below) in three-player games by guiding the search with two values: one that represents the actual goal of winning, and one that is more finegrained, representing some "cardinal score" the win/loss is based on.

Critical strengths

None.

Critical weaknesses

First, the paper is written in such dense Claude jargon that it is basically unreadable.

Examples from the abstract alone:

  • "Objectives such as win probability or rank credit make that aligned value coarse: a player’s actions are mostly exactly tied or far apart, so a few-simulation search over the aligned value is often left undecided." This is very surprising to me. What kind of search is assumed here? Even in MCTS with random rollouts, I have rarely seen two moves get exactly the same evaluation.
  • "also backs up the cardinal head’s vector" So there is a network head? But then there would not be undecided searches, even at low simulations. What else could be meant by "head"? Similarly, what is a "readout"? I've never seen this word in the hundreds of search papers I've read.
  • "where the aligned margin clears a calibrated gate threshold" is typical Claude word salad - what does this mean? Does it for example mean, "where the difference in winrate between the best and second-best move is larger than a hyperparameter"? The reader doesn't know.
  • "On an exactly solved eight-token ranking game and on a reduced threeplayer Chinese Checkers, the gated readout at 64 simulations per move" Is this using MCTS? What is it guided by?

Continuing in the introduction:

  • What are "the Tally probe states"?
  • "On the same game [what game?] and recipe [what recipe?] at two players, the aligned readout’s [what is a readout? simply the average winrate estimate?] regret [what regret?] on the win network [what is the win network?] is less than a third of its three-player value [what three-player value?]" This is typical for Claude, which struggles after a long interaction to understand the difference between the human it has just discussed the problem and written the paper with, and the future reader of the paper who does not have access to the same conversation history.
  • What does "at the cost of one search" mean? How do you lose "a search"?
  • "On the win network that action is rank-dominated often enough to be a worse decision than the aligned readout." So a readout seems to mean a chosen action. What does it now mean that an action is "rank-dominated" by another "on the win network"? Are we talking about averages of winrate estimates at the root? And rank-dominated means "less than"?
  • "48 token tables, each a different game, with three network seeds per table" What is a "token table", some randomized starting position of this game? How do network seeds matter for the computation of exact game-theoretic values?
  • What's an "endgame band"? The last k moves of the game for some k?
  • "In full games against two copies of a fixed AlphaZero-style opponent (...) [on] both training objectives" - Why is the paper first computing exact values, and then measuring performance against an AlphaZero opponent instead? Why only one, do the results generalize? And what two training objectives?
  • "It also beats the cardinal search at 64 and 128 simulations" - Does this now mean beating it playing against each other, or still playing against a third opponent? etc. etc.

Second, the paper also turns out to propose an incremental, situational improvement to an algorithm that doesn't really exist in the literature - for good reason, I believe. The text mostly hides this, using names for the algorithm ("vector maxn search", "PUCT with the maxn backup") that sound like well-known baselines (MCTS-max^n); only one sentence hints at this unknown algorithm: In Section 2, "The maxn backup replaces a node’s vector by that of the child maximizing the mover’s component" I only know what algorithm the paper is actually about because the authors, to their credit, made their codebase available. The paper is about a best-first kind of PUCT where a node’s vector is replaced by that of the best child, with no averaging. There is nothing like this in the reference list, which the paper does not mention. Instead, the paper presents the replacing backup as if it were the standard one. Section 2 attributes “the maxn backup” to Luckhardt & Irani, and the related-work section says per-player vectors “backed up per player inside an AlphaZero loop appear in Petosa & Balch (2019).” A reader would think that the search follows existing multi-player AlphaZero practice. It doesn’t: Petosa & Balch average. The code repeats the impression, describing its search as “the natural MuZero-style MCTS adapted to maxn.” The natural adaptation in the literature, MCTS-maxn as Nijssen and Petosa & Balch use the term, is the averaging one.

Questions for authors

None.

12 comments

Dear Reviewer c8aV,

Thank you for the time you spent on our submission and for the detailed review. Your comments identified passages that were difficult to follow and an attribution of our backup rule that was not clear, and we have revised the manuscript on both points. We respond to each concern below.

Comment 1. AI use and the authors' contribution

The authors are not native English speakers, and we used Claude to help draft, translate and edit the manuscript. We regret that the resulting text was difficult to read. The other uses of AI tools are disclosed in the AI use statement of the submission.

The research was defined and led by the authors. We set the research question and proposed the central design: independently backing up the objective-aligned value and the cardinal score in one tree, then using the gap between the two highest aligned-value estimates to choose the final action. We selected the games, designed the evaluation combining exact decision regret with full-game performance, and directed every experiment on our own hardware. We reviewed the proofs and the code, and checked the reported numbers against the result files and the references against their published records.

The results can be verified independently of the text. Every table and figure is generated by a script from result files included with the submission, every experiment is a deterministic function of recorded seeds, and a test suite checks the searches against the exact solver.

To address the readability concern, we have rewritten the abstract and the introduction so that each term is defined before it is used. We would be glad to answer any further questions about the method, the experiments or the authors' role in them.

Comment 2. Method, attribution and evidence

W2a. The replacement backup and its relation to prior work.

Thank you for examining the code. You are right that our search does not use the mean backup of Petosa & Balch, and that the submitted text did not state this difference clearly. The revised introduction, §2 and §6 now distinguish the classical max⁡n\max^n operator (Luckhardt & Irani), its use as a replacement backup in our partially expanded PUCT tree, and the mean backup of multiplayer AlphaZero (Petosa & Balch).

The submitted §2 described the searches as “per-player-vector variants of PUCT” and stated that the backup replaces a node's vector with the child vector maximizing the mover's component, but it did not contrast this rule with the mean backup. We have also corrected the code comment that described the implementation as a “natural” adaptation; the search behavior is unchanged.

W2b. The contribution and whether it is confined to the original search.

The improvement is not confined to the replacement-backup search: adding the same gate to mean-backup PUCT improves it in all four game–objective settings. We first restate the method and then give this comparison.

Our method has three steps:

  1. Grow a search tree using the policy prior and the value aligned with the win or rank objective.
  2. Independently back up the cardinal score vector through the same visited nodes, leaving tree selection and expansion unchanged.
  3. At the root, use the cardinal value to choose the final action when the gap between the two highest aligned-value estimates among visited root actions is below the threshold; otherwise retain the base search's recommendation.

Our contribution connects measurements of how the two decisions' errors vary with the aligned margin to this final-action rule (§3, Figure 2). The controls separate the effects of final action selection, tree guidance and selective switching (§5.4).

We directly tested the construction on mean-backup PUCT (Mean). Mean+gate keeps Mean's selection, expansion, visit counts and aligned mean statistics, and adds the cardinal vector, backed up by replacement as in §4, and the same gate. With the 300-game networks and opponents fixed, both methods use 64 simulations. The networks and opponents were trained with replacement backup. Table 13 (Appendix E) reports paired tournament-credit differences against the same two copies of a fixed win-trained opponent, with pointwise 95% bootstrap intervals over tables or openings. CC denotes Chinese Checkers.

ComparisonTally rankTally winCC rankCC win
Mean+gate − Mean+.006 [.00009,.012]+.005 [.001,.009]+.229 [.205,.254]+.313 [.276,.346]

All four intervals are above zero. Tables 2 and 13 and Appendix L give the additional search and training-level comparisons.

Comment 3. Presentation (1/2)

W1. Responses to the passages cited in the review.

Thank you for listing the specific passages; they showed us which terms and comparisons the submitted text had left undefined. We address each example in the order it appears in the review. Items 7–12 follow in the next comment.

  1. “actions are mostly exactly tied or far apart”; “search … left undecided.”

    Exact ties refer to solved game values under the win or rank payoff, not to equality of network predictions or sampled return averages. Our PUCT search uses a learned policy and aligned value; at low budgets it can choose an inferior action even when all estimates differ. By “undecided” we meant an action selection that is susceptible to small estimation errors. The revised text no longer uses this word and separates exact ties from estimated values and decision errors.

  2. “the cardinal head's vector”; “readout.”

    A head is a network output: the aligned head predicts each player's win or rank credit, and the cardinal head predicts each player's raw score. A readout is the final action-selection rule. The aligned readout chooses the most-visited root action; the cardinal readout chooses the visited root action with the largest backed-up cardinal value for the player to move.

  3. “the aligned margin clears a calibrated gate threshold.”

    The aligned margin is the difference between the two highest aligned-value estimates among visited root actions, for the player to move. At or above a threshold calibrated offline using exact action values, we play the most-visited root action; below it, we play the visited action with the largest backed-up cardinal value.

  4. “the gated readout at 64 simulations per move”: what search and guidance?

    This is our PUCT variant with 64 simulations per move, guided by the policy prior and aligned head and using the disclosed replacement backup. The cardinal head is backed up independently through the same visited nodes and is consulted only for final action selection.

  5. “the Tally probe states.”

    These are reachable positions in the eight-token Tally game where the player to move has legal actions with different exact objective values. We sample up to twelve per token table and objective and evaluate chosen actions with the exact solver.

  6. “On the same game and recipe at two players … regret … on the win network.”

    This compares two- and three-player Tally instances with the same training recipe and 64 simulations per move. A win network is trained to predict win credit. We measure the exact regret of its most-visited root action: the best exact win value minus the chosen action's exact win value, assuming subsequent play follows the exact solution under a fixed tie convention. The means are 0.047 at two players and 0.161 at three players (Appendix D.1).

Reviewer_WT5HOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes to use winning rate when the best move (e.g. the most simulated) is sufficiently above the second best move and to use the score rate when it is not the case. It has good results in Chinese Checkers and Tally.

Critical strengths
  • The algorithm improves on Alpha Zero for multiplayer games
  • The idea is simple and concise
  • The paper is well written
Critical weaknesses
  • There is on hyperparameter to tune (the Delta for the difference in winning rates)
  • Other more difficult multiplayer games should be included in the experiments
Questions for authors
  • Did you try other more complex games?
  • How sensitive to Delta is the algorithm?
1 comment

Dear Reviewer WT5H,

Thank you for your careful reading of our paper and for the positive assessment. Your two questions concern the sensitivity of the method to its threshold and its behavior in more complex games. We answer each below and give the corresponding locations in the revised manuscript.

Q2 / W1. How sensitive is the method to the threshold Δ\Delta?

Halving or doubling the threshold preserves the improvement over the same search without the gate in all four game–objective settings, with every 95% interval above zero. The threshold is cΔ^Ac\hat{\Delta}_A: c=2c=2 is shared across games, while Δ^A\hat{\Delta}_A is calibrated per game and objective using exact action values and a zero-value search. Both constants were fixed before the tournaments, without using trained networks or tournament outcomes.

Table 14 gives gated minus aligned tournament credit at 64 simulations against the same two copies of a fixed win-trained opponent. Intervals are 95% bootstrap intervals over tables or openings; parentheses give firing rates, and bold marks intervals excluding zero. CC denotes Chinese Checkers.

ThresholdTally rankTally winCC rankCC win
c=1c=1 (half)+.009 [.004,.013] (0.550)+.013 [.007,.019] (0.577)+.262 [.240,.283] (0.973)+.354 [.316,.392] (0.956)
c=2c=2 (paper)+.012 [.006,.018] (0.705)+.017 [.010,.024] (0.703)+.263 [.241,.285] (0.991)+.363 [.325,.400] (0.986)
c=4c=4 (double)+.010 [.004,.017] (0.852)+.021 [.013,.029] (0.817)+.264 [.242,.286] (1.000)+.367 [.331,.402] (0.994)
Other game's error scale, c=2c=2+.002 [-.006,.009] (0.982)+.019 [.011,.027] (0.868)+.261 [.241,.281] (0.906)+.347 [.309,.384] (0.899)

The Tally firing rate changes from about 55% to 85% across c=1,2,4c=1,2,4, while the improvement persists. CC favors the cardinal readout across these settings, with firing rates above 95%. Using the other game's error scale preserves a resolved gain in three of four settings; the Tally rank interval includes zero, supporting our use of game-specific calibration.

Manuscript locations: §2, §5.4 and Table 14 (Appendix E).

Q1 / W2. Have you tried more complex multiplayer games?

Yes. We tested the method on three-player Rolit, and with networks trained on 300 self-play games the gated readout improved on the same search without the gate under both objectives.

Rolit is a multiplayer variant of Othello on the standard 8×8 board that has been used as a test game for multiplayer tree search (Nijssen, 2013). Every game lasts 60 moves with an average branching factor of about 8.5, and the three-player state space is bounded by about 103810^{38} (Nijssen, 2013), so exact values are available only near the end of a game. We kept the networks, training recipe, opponent and tournament protocol of the paper. The threshold was not tuned for this game: c=2c=2 as in the paper, and the error scale was calibrated by the paper's procedure on solvable endgame states before any network was trained. The comparisons were registered in advance.

The table gives the tournament credit of the gated readout at 64 simulations minus the same search without the gate, against two copies of a fixed 300-game win-trained opponent, with 95% bootstrap intervals over 64 openings. Bold marks intervals excluding zero.

Gated at 64 simulations minusRankWin
Same search without the gate, 64 simulations+.174 [.145,.203]+.254 [.216,.293]
Same search without the gate, 128 simulations+.109 [.077,.140]+.204 [.169,.238]

The gated readout at 64 simulations also exceeds the same search without the gate at twice the budget. As in Chinese Checkers, the aligned margins of the trained networks are mostly below the threshold, so the gate fires on 99 percent (rank) and 88 percent (win) of moves, and the gain comes mainly from the cardinal readout of the aligned-guided tree.

We thank the reviewer for this suggestion, which led us to test the method beyond Tally and Chinese Checkers. The Rolit result suggests that the method may extend to more complex games, and we report it in Appendix M as an additional study. This study covers one board size and the small networks of the paper. It does not repeat the comparisons with other search methods or the controls of §5.3–§5.4. Because exact values are available only near the end of a game, its position-level results are limited to the exact regret on 32 solved endgame states (Table 35) and do not include the margin-band analysis of §3. A full evaluation on larger games remains future work.

Manuscript location: Appendix M (Table 35).

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.