PaperlayerView exact revision on arXiv

Source-linked AI summary

Winning Before You Play: An Iterative Capture Algorithm for Pre-Game Dominance Analysis in Three-Player Auction Bridge

Sourish Sarkar

arXiv:2609.29615v1cs.GT

TL;DR

Traditional four-player auction bridge leaves the winning bidder’s partner passive; this paper proposes a three-player variant and a structural capture estimator. Among qualifying deals, over 93% of capture totals fall between 8 and 11.

  • Problem

    Traditional four-player variants have a passive partner role, motivating a three-player auction-bridge design.

    Source: §I · p. 1

  • Method

    The paper uses a closed-form iterative capture estimator that derives trick estimates from the dealt hand’s structure rather than learned models or game-tree search.

    Source: §IX · p. 9

  • Results

    Over 93% of qualifying outcomes have total captures T ∈{8, 9, 10, 11}, with the distribution stable across independent runs.

    Source: D. Discussion · p. 6

  • Takeaways & Limitations

    The estimator provides an exact, reproducible trick estimate in time independent of the number of simulated trials.

    Source: §IX · p. 9

    The greedy smallest-gap-first rule is a heuristic for randomly dealt hands and is not expected to be optimal against an opponent trying to beat it.

    Source: Where Reinforcement Learning Could… · p. 9

Abstract

from arXiv · show

This paper introduces an innovative and bias-free variant of a three-player auction bridge, specifically designed to address structural limitations found in traditional formats. In a standard auction bridge, the partner of the winning bidder often plays a passive role, possessing minimal influence over the outcome once the bidding phase concludes. To rectify this, our proposed three-player model redistributes strategic agency, emphasising a framework where No Trump strategies statistically outperform specific trump card selections. The core of this research is a rule-based dual-algorithm framework. First, we implement an iterative bidding algorithm that allows players to calibrate their bids dynamically based on evolving hand strengths. Second, we introduce a predictive iterative algorithm designed to calculate the expected number of tricks a player will secure before the lead card is played. Unlike traditional approaches that rely on backtracking or high-complexity dynamic programming, our iterative method significantly reduces time complexity. This efficiency is crucial for real-time gameplay, enabling players to perform complex probability assessments instantly without computational lag. To validate the model, a simulation was conducted on a dataset of 1,00,000 game instances. This large-scale analysis allowed for a precise comparison between initial bidding, outcomes, and overall bidding accuracy. The empirical results demonstrate that our algorithm not only enhances the balance and fairness of the auction bridge but also exhibits high reliability in predictive accuracy. Beyond the realm of card games. By modelling uncertainty and strategic competition, the algorithm offers valuable insights into risk management and decision-making in volatile market environments.

I. INTRODUCTION

The paper proposes a three-player auction-bridge variant to address the passive partner role in traditional four-player formats and introduces a structural alternative to search-based capture estimation.

I. INTRODUCTION

  • Prior bridge systems use sampling and search, learned bidding –, or interpretable rule induction; this paper instead uses structural, constant-time conditions.Source: §I · p. 1

A. Core Mechanics and Rule Framework

A. Overview

  • The estimator represents a standard deck as a 4 × 13 matrix and computes suit captures from the deal’s structure without simulating gameplay.Source: A. Overview · p. 2

B. Definitions and Notation

The section defines suit-level counts and budgets used by the capture procedure, then constructs a matrix encoding each dealt card’s owner.

B. Definitions and Notation

  • For each suit, c1(s), c2(s), and c0(s) count Player 1’s cards, Player 2’s cards, and empty positions, respectively.Source: B. Definitions and Notation · p. 2
  • The leading gap δ(s) is the first nonzero column index, or 13 for an all-zero row; Δ totals these gaps across suits.Source: B. Definitions and Notation · p. 2
  • Z is the global zero budget, while Zs is the per-suit budget; M(s) limits captures, and κ(s), π(s), and O(s) track captures, card position, and owned positions.Source: B. Definitions and Notation · p. 2

C. Step 1 — Construct the State Matrix

D. Step 2 — Row Analysis

The procedure analyzes each suit, rejects infeasible deals before capture, and updates the deal after captures by removing an opponent card.

D. Step 2 — Row Analysis

  • Row analysis computes each suit’s leading gap, player counts, empty slots, and maximum captures, then derives total gap Δ and budget Z.Source: D. Step 2 — Row Analysis · p. 2

E. Step 3 — Pre-Qualification

F. Step 4 — Opponent Replacement Sub-Procedure

G. Step 5 — Initial Captures (One Per Suit)

H. Step 6 — Main Capture Loop (Smallest Gap First)

The main loop repeatedly checks for a stop condition, selects an affordable suit by gap, and performs a capture before updating the state.

H. Step 6 — Main Capture Loop (Smallest Gap First)

6.2 — Select Best Suit

  • For each suit, the algorithm computes raw and effective gaps, then selects the best valid suit; it stops if no valid suit exists.Source: §6.2 · p. 3
  • If the smallest available gap exceeds the global budget while the selected suit’s row budget remains positive, the loop stops.Source: §6.2 · p. 3

6.4 — Perform the Capture

  • For the selected suit, the procedure advances its pointer and capture counts, accounts for budget availability, replaces an opponent card, and returns to the stop check.Source: §6.4 · p. 3

I. Step 7 — Final Output

Upon loop termination, the algorithm reports its final state matrix, total captures, and final check value; the variable glossary names the quantities used to track captures and suit gaps.

I. Step 7 — Final Output

  • At termination, the algorithm outputs the final state matrix D, total captures, and final check value.The matrix reflects opponent replacements applied during captures.Source: §I · p. 3

J. Summary of Variables

  • The variable glossary defines per-suit zero budgets, capture limits, nonzero-card indices, raw and effective gaps, and accumulated captures.The effective gap is zero when a row budget is exhausted; otherwise it equals the raw gap.Source: J. Summary of Variables · p. 4

III. PREDICTING SUIT DOMINANCE: A COMBINATORIAL CARD CAPTURE ALGORITHM

The estimator uses the structural distribution of two players’ 26 cards across four suits to count achievable suit captures without simulating gameplay. It initializes a state matrix, checks qualification conditions, then applies initial and greedy captures.

Problem Statement

  • The estimator determines the total suit captures achievable from a random deal of 26 cards to two players without simulating gameplay.Its input is the initial card distribution in a standard 52-card deck.Source: Problem Statement · p. 4
  • The algorithm represents card ownership in a 4×13 state matrix and computes each suit’s leading gap, card counts, capture limit, zero budget, and ordered occupied indices.It sets the global zero budget to six minus the total leading gap and initializes the capture count to zero.Source: Notation · p. 4, Algorithm · p. 4
  • It returns zero captures if the total leading gap exceeds six or either player’s count in any suit is no greater than that suit’s leading gap.Otherwise, it begins with one free capture per suit and removes the opponent’s rightmost card in that suit.Source: Algorithm · p. 4
  • The greedy loop selects the eligible suit with the smallest effective gap, captures there, updates budgets and counters, and replaces the opponent’s rightmost card.It stops when the global condition is met, no eligible suit remains, or the next gap exceeds the available budget.Source: Algorithm · p. 4

Complexity Analysis

The analysis describes the estimator as constant-time and constant-space for a standard deck, and gives a polynomial worst-case bound for a generalized deck with n suits and m ranks.

Complexity Analysis

  • O(52) is the stated complexity for initial captures; the listed stages also include initialization, pre-qualification, and scanning suits to compute gaps, select a suit, and replace a card.These stage descriptions are presented for a standard deck.Source: Complexity Analysis · p. 5
  • O(1) time and O(52) = O(1) space describe the fixed-deck estimator, which the paper contrasts with iterative MCTS/UCT, and regret-minimisation methods.The paper attributes the bound to the fixed four-suit, 13-rank domain and the state matrix D.Source: Complexity Analysis · p. 5
  • O(n^2m^2) is the stated worst-case time complexity for a generalized deck with n suits and m ranks.The analysis gives O(n · m) for initialization and per greedy-loop iteration, with at most n · m iterations.Source: Complexity Analysis · p. 5

IV. SIMULATION STUDY

Across three runs of one million random deals, the study evaluates pre-qualification and capture outcomes; qualifying deals cluster in mid-range captures, with none exceeding the algorithm’s bound.

A. Experimental Setup

  • The simulation applies the algorithm to uniformly random 13-card hands dealt to two players, recording capture counts over 1,000,000 trials per run and three runs.The authors frame this as a two-player zero-sum simulation.Source: A. Experimental Setup · p. 5

C. Figures

  • Figures 1–2 show pre-qualification outcomes and qualifying-trial capture distributions; run-to-run overlap indicates stable patterns.Figure 1 displays grouped trial-count bars and a mean-percentage line; Figure 2 displays the capture-count distribution.Source: C. Figures · p. 5, Figure 1 · p. 5, Figure 2 · p. 6

B. Results

  • T=9 and T=10 jointly account for roughly 18.3% of all trials (about 65.6% of qualifying trials), while T≤6 occurs in fewer than 0.19% of qualifying trials.Table I includes only trials satisfying both pre-qualification conditions.Source: B. Results · p. 5, Table I · p. 6
  • Only 27.90% of deals pass both conditions, while 57.57% satisfy Condition A but fail Condition B.The pre-qualification proportions vary by less than 0.1 percentage points across runs.Source: B. Results · p. 5, Table II · p. 6

D. Discussion

  • Among qualifying deals, T∈{8, 9, 10, 11} accounts for over 93%; no T>13 occurred in 3,000,000 trials, supporting the stop condition.The capture distribution was stable across independent runs.Source: D. Discussion · p. 6

V. COMPARISON BETWEEN 13-CARD SINGLE HAND BIDDING ALGORITHM VS FULL HAND FINAL OUTPUT

The comparison evaluates how revealing the dummy’s hand changes pre-bid trick estimates and how second-hand criteria affect qualification. Three independent simulation runs are used to assess whether the reported results are stable.

V. COMPARISON BETWEEN 13-CARD SINGLE HAND BIDDING ALGORITHM VS FULL HAND FINAL OUTPUT

  • Revealing the dummy’s hand changes the trick estimate, providing a data-driven basis for bidding strategy in Auction Bridge.The comparison contrasts a pre-bid estimate based on one 13-card hand with a revised estimate using the combined 26-card layout.Source: §V · p. 7

VI. IMPORTANCE OF 2ND HAND CRITERIA AND SIMULATION ANALYSIS

  • About 25–28% of cases that pass Algorithm 1 are subsequently rejected because the second hand fails Condition B.Condition B fails when Player 2’s count in any suit is less than or equal to that suit’s leading gap.Source: §VI · p. 7

VII. WHY WE RAN THE SAME SIMULATION THREE TIMES

  • The three independent simulation runs agree almost to the decimal point, supporting the stability of their reported results.The runs use fresh, independent shuffles.Source: §VII · p. 7

VIII. THE THREE RUNS, SIDE BY SIDE

  • Table IV summarizes three 10-million-deal runs, including No-Trump qualification, contract outcomes, and scoring under both schemes.Each run shuffles and deals 10 million hands.Source: §VIII · p. 7

The story the numbers tell

Across three independent runs, contracts are made more often when both the bidder and dummy independently qualify than in the overall set of opened contracts. Most losses occur when the dummy does not qualify, while the scoring scheme changes the scale of net points; the paper concludes with a constant-time structural capture estimator.

The story the numbers tell

  • 99.8% or more of losses occur when the dummy’s hand did not independently qualify, across all three runs.The paper attributes the shortfall to the unknown strength of the face-down partner’s hand, rather than a failure to resolve the capture sequence.Source: The story the numbers tell · p. 8
  • The new scoring scheme yields roughly 819–833 thousand net points, compared with 152–155 thousand under the old scheme.The paper says the new scheme rewards overtricks more generously; run ordering is similar but not identical.Source: The story the numbers tell · p. 8, Figure 6 · p. 9

A visual summary

  • Figure 6 places win-rate comparisons, net points under both scoring schemes, and both-eligible deal counts and losses side by side.The both-eligible subgroup has over ten thousand cases in each run.Source: A visual summary · p. 9, Figure 6 · p. 9

The story the numbers tell

  • 93.42%, 93.50%, and 93.75% of both-eligible contracts are made, versus 65.86%, 66.00%, and 66.03% of opened contracts overall.Figure 6 compares overall and both-eligible win rates across the three runs.Source: The story the numbers tell · p. 8, Figure 6 · p. 9

Why the agreement between runs matters

  • Across three independent samples of ten million deals, overall win rates differ by less than 0.2 percentage points and both-eligible win rates by less than 0.4 points.The paper interprets this agreement as evidence that the simulation has converged.Source: Why the agreement between runs matters · p. 9

IX. CONCLUSION

  • The paper presents a constant-time structural estimator for trick-taking potential, while leaving competitive bidding among three active players as a future extension.The estimator is studied in a two-player zero-sum capture setting.Source: §IX · p. 9

Where Reinforcement Learning Could Take This Next

The section sketches reinforcement-learning extensions for sequential bidding and post-dummy card play, alongside Bayesian updating of opponent-hand beliefs. It proposes that evolving beliefs could inform both capture estimates and learned policies.

Where Reinforcement Learning Could Take This Next

  • The estimator is rule-based and closed-form; its greedy, smallest-gap-first heuristic is motivated for random deals, with no reason to expect optimality against an actively adversarial opponent.Reinforcement learning is proposed as a layer built on top of the structural estimator, not as its replacement.Source: Where Reinforcement Learning Could… · p. 9
  • Bidding could be modeled as a sequential policy over hand and bid history, with bid-or-pass actions and round-score rewards; three independent bidders distinguish this variant from fixed partnerships.Prior work trained bidding policies for standard four-player bridge, while this variant's multi-agent setting is described as largely unexplored.Source: Where Reinforcement Learning Could… · p. 9
  • After the dummy is revealed, self-play training could produce a card-play policy for head-to-head comparison with the greedy rule, measuring the strategic ceiling the deterministic algorithm leaves unknown.The proposed self-play direction is linked to AlphaGo and counterfactual regret minimisation methods used in superhuman poker play.Source: Where Reinforcement Learning Could… · pp. 9–10

Where Bayesian Estimation Could Take This Next

  • Bayesian estimation would update opponent-hand beliefs from random-deal priors, played cards, and bids, then use posterior sampling to approximate card-win probabilities.Failing to follow suit signals zero cards in that suit; the text proposes sampling plausible hands because exact summation is too large.Source: Where Bayesian Estimation Could Take… · p. 10
  • Under this extension, c1(s), c2(s), and δ(s) become random variables and total captures T becomes an expected value, with opponent-hand beliefs also serving as inputs to learned policies.The proposed belief-based estimator retains the algorithm's underlying structure.Source: Where Bayesian Estimation Could Take… · p. 10
Loading 2609.29615v1…