The Gold Standard of Zero-Sum Games | Separating Exploitability from Observed Win Rate

Exploitability is defined relative to a game, payoff convention and strategy space. Write the row player’s payoff matrix and remember that the column player’s payoff is its negative in a two-player zero-sum game. A win count alone discards payoff magnitude and game value.

Begin with a measurement ledger. Name the quantity, its unit, the sample or observation that produced it, the time window and the transformation rule. Mark each input as observed, stipulated, estimated or assumed. This prevents a familiar label from importing a new denominator or a stronger claim. Read the ledger as a sentence before calculating; if the sentence is incomplete, the arithmetic is premature.

Use the fictional row-payoff matrix with actions A and B against columns L and R: A,L=1; A,R=−1; B,L=−2; B,R=2. Positive values favour row, negative values favour column. This is a teaching game, not a model of a real competitive setting.

Before moving on, record the observation that would distinguish the two most plausible explanations. If no feasible observation exists, narrow the claim instead of pretending the ambiguity has disappeared.

A changed-context transfer check is to add 5 to every matrix entry and explain why the game is no longer represented by the same zero-sum value convention unless the opposing payoffs and baseline are adjusted consistently. State what would change the result and what the example cannot establish. That limit belongs inside the answer.

Contents · Next chapter

For a 2×2 mixed equilibrium, choose probabilities that make the opponent indifferent when no pure saddle point exists. The game value is the expected payoff secured by equilibrium play, not the average from any one sample.

Keep the mechanism beside the number. Ask what physical, biological, statistical or strategic relationship connects the observation to the conclusion, and which conditions could break that link. A correct calculation can answer the wrong question when a changed setting alters the process. Draw one arrow for every inferential step, then challenge each arrow with a counterexample.

Let row play A with probability p. Against L the expected row payoff is p(1)+(1−p)(−2)=3p−2. Against R it is p(−1)+(1−p)2=2−3p. Equating them gives p=2/3 and value 0.

Before moving on, record the observation that would distinguish the two most plausible explanations. If no feasible observation exists, narrow the claim instead of pretending the ambiguity has disappeared.

A changed-context transfer check is to substitute p=2/3 into both expressions and verify the common value before using it as the baseline. State what would change the result and what the example cannot establish. That limit belongs inside the answer.

Contents · Previous chapter · Next chapter

The column mixture should make row indifferent. Keeping both sides avoids the common error of finding one player’s probability and assuming the other must match it.

Use an independent check that follows a different route. Reconstruct totals from components, compare a bound with a point estimate, inspect a distribution behind an average, or solve a changed case with the original definitions. Agreement does not prove every assumption, but disagreement locates a definition, unit or boundary that needs repair.

Let column play L with probability q. Row’s payoff from A is q(1)+(1−q)(−1)=2q−1. From B it is q(−2)+(1−q)2=2−4q. Equality gives q=1/2 and value 0.

Before moving on, record the observation that would distinguish the two most plausible explanations. If no feasible observation exists, narrow the claim instead of pretending the ambiguity has disappeared.

A changed-context transfer check is to change only the A,L payoff and solve both mixtures again rather than adjusting one probability by intuition. State what would change the result and what the example cannot establish. That limit belongs inside the answer.

Contents · Previous chapter · Next chapter

For the row strategy, exploitability in this teaching matrix is the game value minus the worst payoff a best-response column player can force. A strategy at equilibrium has gap zero; a vulnerable strategy has a positive gap.

Match the claim to the evidence. A tidy decimal cannot repair selective observation, and one convenient run cannot establish every product, participant, signal or opponent. Report the result, the conditions that make it interpretable, and the next observation needed for the actual decision. Separate the numerical conclusion from the practical conclusion.

If row uses A with probability 0.8, its expected payoff against L is 3(0.8)−2=0.4, while against R it is 2−3(0.8)=−0.4. Column chooses R, so row’s worst-case payoff is −0.4. Relative to value 0, the row exploitability is 0.4 payoff unit.

Before moving on, record the observation that would distinguish the two most plausible explanations. If no feasible observation exists, narrow the claim instead of pretending the ambiguity has disappeared.

A changed-context transfer check is to evaluate p=0.6 and compare its best-response gap with p=0.8. State what would change the result and what the example cannot establish. That limit belongs inside the answer.

Contents · Previous chapter · Next chapter

Observed win rate belongs to realised actions and opponents. It can be useful for describing that sample but cannot substitute for a best-response calculation.

Begin with a measurement ledger. Name the quantity, its unit, the sample or observation that produced it, the time window and the transformation rule. Mark each input as observed, stipulated, estimated or assumed. This prevents a familiar label from importing a new denominator or a stronger claim. Read the ledger as a sentence before calculating; if the sentence is incomplete, the arithmetic is premature.

Suppose the opponent chooses L in all ten observed games. Row’s randomised play realises A eight times and B twice. The eight A,L outcomes have payoff +1 and the two B,L outcomes −2. Row wins 8 of 10 by payoff sign, an 80 per cent observed win rate, with mean payoff (8−4)/10=0.4.

Before moving on, record the observation that would distinguish the two most plausible explanations. If no feasible observation exists, narrow the claim instead of pretending the ambiguity has disappeared.

A changed-context transfer check is to keep the realised row actions but change every opponent action to R. Recalculate win rate and mean payoff. State what would change the result and what the example cannot establish. That limit belongs inside the answer.

Contents · Previous chapter · Next chapter

A best-response test need not use the opponent seen in the sample. It selects the legal response that minimises the row strategy’s expected payoff.

Keep the mechanism beside the number. Ask what physical, biological, statistical or strategic relationship connects the observation to the conclusion, and which conditions could break that link. A correct calculation can answer the wrong question when a changed setting alters the process. Draw one arrow for every inferential step, then challenge each arrow with a counterexample.

Against the realised 80/20 A/B frequency, column R produces eight payoffs of −1 and two of +2, total −4 and mean −0.4. Thus the same strategy associated with an 80 per cent win rate against L is vulnerable to R. The win record did not measure robustness.

Before moving on, record the observation that would distinguish the two most plausible explanations. If no feasible observation exists, narrow the claim instead of pretending the ambiguity has disappeared.

A changed-context transfer check is to allow column to mix L and R and show why a linear expected payoff is minimised at one of the pure endpoints here. State what would change the result and what the example cannot establish. That limit belongs inside the answer.

Contents · Previous chapter · Next chapter

Win rate counts signs or declared game outcomes; expected payoff also reflects magnitudes. Two records can share a win rate and have different values. Define draws and zero payoffs before calculating.

Use an independent check that follows a different route. Reconstruct totals from components, compare a bound with a point estimate, inspect a distribution behind an average, or solve a changed case with the original definitions. Agreement does not prove every assumption, but disagreement locates a definition, unit or boundary that needs repair.

A record with eight wins of +1 and two losses of −2 has mean +0.4. Another with eight wins of +0.1 and two losses of −2 has mean −0.32 despite the same 80 per cent win rate. The strategic objective must specify which payoff matters.

Before moving on, record the observation that would distinguish the two most plausible explanations. If no feasible observation exists, narrow the claim instead of pretending the ambiguity has disappeared.

A changed-context transfer check is to replace one large loss with several smaller losses and preserve the same total payoff. Compare win rate and mean. State what would change the result and what the example cannot establish. That limit belongs inside the answer.

Contents · Previous chapter · Next chapter

A declared mixed strategy is a probability distribution; ten sampled actions are one realisation. The observed 8/2 split does not prove the intended probability was 0.8. Sampling variation and adaptive behaviour remain possible.

Match the claim to the evidence. A tidy decimal cannot repair selective observation, and one convenient run cannot establish every product, participant, signal or opponent. Report the result, the conditions that make it interpretable, and the next observation needed for the actual decision. Separate the numerical conclusion from the practical conclusion.

An equilibrium strategy with p=2/3 can realise eight A actions in ten trials with nonzero probability. Evaluating the empirical frequency and evaluating the declared policy are different tasks. Record which object the exploitability calculation uses.

Before moving on, record the observation that would distinguish the two most plausible explanations. If no feasible observation exists, narrow the claim instead of pretending the ambiguity has disappeared.

A changed-context transfer check is to increase the sample to one hundred while holding the 80/20 split. Explain what becomes more stable and what selection bias can remain. State what would change the result and what the example cannot establish. That limit belongs inside the answer.

Contents · Previous chapter · Next chapter

Results against a cooperative, fixed or narrow opponent set can overstate robustness. State how opponents were chosen and whether they adapted. Self-play performance, tournament win rate and worst-case response answer different questions.

Begin with a measurement ledger. Name the quantity, its unit, the sample or observation that produced it, the time window and the transformation rule. Mark each input as observed, stipulated, estimated or assumed. This prevents a familiar label from importing a new denominator or a stronger claim. Read the ledger as a sentence before calculating; if the sentence is incomplete, the arithmetic is premature.

If every evaluation opponent always chooses L, the 0.8 strategy looks strong. Adding one best-response policy R changes the strategic picture immediately. This is distribution shift within legal play, not an invalid test.

Before moving on, record the observation that would distinguish the two most plausible explanations. If no feasible observation exists, narrow the claim instead of pretending the ambiguity has disappeared.

A changed-context transfer check is to evaluate against a 50/50 opponent mixture and compare sample performance with worst-case performance. State what would change the result and what the example cannot establish. That limit belongs inside the answer.

Contents · Previous chapter · Next chapter

A response function plots performance against opponents of varying strength or adaptation. It reveals where a strategy succeeds and where it fails. Approximate best responses can be useful in large games, but their limits should be explicit.

Keep the mechanism beside the number. Ask what physical, biological, statistical or strategic relationship connects the observation to the conclusion, and which conditions could break that link. A correct calculation can answer the wrong question when a changed setting alters the process. Draw one arrow for every inferential step, then challenge each arrow with a counterexample.

For the 2×2 teaching game, let column choose L with probability q. Row p=0.8 has expected payoff q(0.4)+(1−q)(−0.4)=0.8q−0.4. The line crosses zero at q=0.5 and reaches its worst value at q=0.

Before moving on, record the observation that would distinguish the two most plausible explanations. If no feasible observation exists, narrow the claim instead of pretending the ambiguity has disappeared.

A changed-context transfer check is to plot the corresponding line for equilibrium p=2/3 and explain why it is flat at zero. State what would change the result and what the example cannot establish. That limit belongs inside the answer.

Contents · Previous chapter · Next chapter

In small matrix games, a best response can be computed exactly. In large sequential games, an evaluation may use an approximate best-response learner. The resulting value can be a lower bound on vulnerability rather than the exact worst-case gap.

Use an independent check that follows a different route. Reconstruct totals from components, compare a bound with a point estimate, inspect a distribution behind an average, or solve a changed case with the original definitions. Agreement does not prove every assumption, but disagreement locates a definition, unit or boundary that needs repair.

The IJCAI paper on approximate exploitability studies learning a best response when exact computation is difficult. Report the algorithm, training budget and evaluation uncertainty instead of writing exploitability as if it were exact.

Before moving on, record the observation that would distinguish the two most plausible explanations. If no feasible observation exists, narrow the claim instead of pretending the ambiguity has disappeared.

A changed-context transfer check is to double the approximate responder’s training budget and decide whether an unchanged result proves optimality. State what would change the result and what the example cannot establish. That limit belongs inside the answer.

Contents · Previous chapter · Next chapter

Different papers and libraries normalise exploitability or NashConv differently. Some sum improvement incentives across players; some average them; some report milli-units per game. Quote the definition before comparing numbers.

Match the claim to the evidence. A tidy decimal cannot repair selective observation, and one convenient run cannot establish every product, participant, signal or opponent. Report the result, the conditions that make it interpretable, and the next observation needed for the actual decision. Separate the numerical conclusion from the practical conclusion.

A two-player profile may have a 0.2 improvement opportunity for row and 0.4 for column. A summed NashConv is 0.6, while an average-style exploitability might report 0.3. Neither label is interpretable without its formula.

Before moving on, record the observation that would distinguish the two most plausible explanations. If no feasible observation exists, narrow the claim instead of pretending the ambiguity has disappeared.

A changed-context transfer check is to convert a reported sum to an average only after confirming the player count and convention. State what would change the result and what the example cannot establish. That limit belongs inside the answer.

Contents · Previous chapter · Next chapter

A best response is a mathematical optimisation within the game. It may be brittle against other opponents or unrealistic as a human style, yet it still exposes a weakness in the tested strategy.

Begin with a measurement ledger. Name the quantity, its unit, the sample or observation that produced it, the time window and the transformation rule. Mark each input as observed, stipulated, estimated or assumed. This prevents a familiar label from importing a new denominator or a stronger claim. Read the ledger as a sentence before calculating; if the sentence is incomplete, the arithmetic is premature.

Column R is the best response to p=0.8 in the teaching matrix even if an observed population rarely chooses R. The exploitability result says R can force −0.4 under the game, not that every future opponent will discover or use it.

Before moving on, record the observation that would distinguish the two most plausible explanations. If no feasible observation exists, narrow the claim instead of pretending the ambiguity has disappeared.

A changed-context transfer check is to restrict the opponent’s allowed strategies and explain why exploitability must be redefined for the smaller space. State what would change the result and what the example cannot establish. That limit belongs inside the answer.

Contents · Previous chapter · Next chapter

A best response may use the strategy distribution and legal information, but it must not see private random choices before acting unless the game permits that. Leaking realised actions creates an impossibly strong evaluator.

Keep the mechanism beside the number. Ask what physical, biological, statistical or strategic relationship connects the observation to the conclusion, and which conditions could break that link. A correct calculation can answer the wrong question when a changed setting alters the process. Draw one arrow for every inferential step, then challenge each arrow with a counterexample.

Column may know that row uses p=0.8 and choose R before row’s random draw. It may not observe the realised A or B and then switch columns if the simultaneous game hides that action. The information structure is part of the game.

Before moving on, record the observation that would distinguish the two most plausible explanations. If no feasible observation exists, narrow the claim instead of pretending the ambiguity has disappeared.

A changed-context transfer check is to turn the simultaneous move into a sequential observed move and recompute the strategic problem. State what would change the result and what the example cannot establish. That limit belongs inside the answer.

Contents · Previous chapter · Next chapter

A strategy tailored to one opponent can outperform equilibrium play against that opponent while becoming more exploitable. This is not automatically bad; it is a trade-off that should be named.

Use an independent check that follows a different route. Reconstruct totals from components, compare a bound with a point estimate, inspect a distribution behind an average, or solve a changed case with the original definitions. Agreement does not prove every assumption, but disagreement locates a definition, unit or boundary that needs repair.

Against always-L, pure A earns 1 per game, better than equilibrium’s expected zero. Against best-response R, pure A earns −1. The gain against the modelled opponent comes with a larger worst-case gap.

Before moving on, record the observation that would distinguish the two most plausible explanations. If no feasible observation exists, narrow the claim instead of pretending the ambiguity has disappeared.

A changed-context transfer check is to mix a small probability of B into pure A and trace both opponent-specific payoff and worst-case payoff. State what would change the result and what the example cannot establish. That limit belongs inside the answer.

Contents · Previous chapter · Next chapter

Repeated games can support uncertainty estimates for observed win rate or mean payoff under a sampling process. They do not turn that sample into exact exploitability. Exploitability uncertainty comes from game estimation, policy estimation or approximate response computation.

Match the claim to the evidence. A tidy decimal cannot repair selective observation, and one convenient run cannot establish every product, participant, signal or opponent. Report the result, the conditions that make it interpretable, and the next observation needed for the actual decision. Separate the numerical conclusion from the practical conclusion.

An 8/10 win record is visibly uncertain as a population win probability estimate. Even with a narrow interval from many games against L, the result would still concern L-like opponents. The best-response gap needs strategic coverage, not only more of the same matchup.

Before moving on, record the observation that would distinguish the two most plausible explanations. If no feasible observation exists, narrow the claim instead of pretending the ambiguity has disappeared.

A changed-context transfer check is to increase the number of games against L to ten thousand and explain why R remains an unresolved opponent. State what would change the result and what the example cannot establish. That limit belongs inside the answer.

Contents · Previous chapter · Next chapter

If a policy changes during evaluation, one pooled win rate may describe no single strategy. Store version and time. An opponent can also adapt, making early and late games incomparable without a model.

Begin with a measurement ledger. Name the quantity, its unit, the sample or observation that produced it, the time window and the transformation rule. Mark each input as observed, stipulated, estimated or assumed. This prevents a familiar label from importing a new denominator or a stronger claim. Read the ledger as a sentence before calculating; if the sentence is incomplete, the arithmetic is premature.

Suppose the first five games use p=1 and the next five use p=0.6. The pooled empirical frequency may be 0.8, but exploitability of a fixed p=0.8 policy is not automatically the exploitability of this switching process.

Before moving on, record the observation that would distinguish the two most plausible explanations. If no feasible observation exists, narrow the claim instead of pretending the ambiguity has disappeared.

A changed-context transfer check is to evaluate each phase separately and then define what a time-aware adversary can observe. State what would change the result and what the example cannot establish. That limit belongs inside the answer.

Contents · Previous chapter · Next chapter

The existing article on one strategy across several payoff scenarios asks whether one choice protects guarantees when the payoff matrix is uncertain. Exploitability here assumes a defined game and varies the opponent response. Do not swap those uncertainties.

Keep the mechanism beside the number. Ask what physical, biological, statistical or strategic relationship connects the observation to the conclusion, and which conditions could break that link. A correct calculation can answer the wrong question when a changed setting alters the process. Draw one arrow for every inferential step, then challenge each arrow with a counterexample.

A policy can be unexploitable in one known matrix yet perform poorly if the true matrix is different. Conversely, a robust policy across matrices can remain exploitable within each. Record whether uncertainty is about payoffs, opponents or policy execution.

Before moving on, record the observation that would distinguish the two most plausible explanations. If no feasible observation exists, narrow the claim instead of pretending the ambiguity has disappeared.

A changed-context transfer check is to change the payoff matrix while holding the opponent fixed and contrast that with a best-response change inside one matrix. State what would change the result and what the example cannot establish. That limit belongs inside the answer.

Contents · Previous chapter · Next chapter

The University of Alberta response-functions paper defines exploitability through best responses and game value. The OpenSpiel research paper describes game-learning evaluations, while the approximate-exploitability paper explains learned responses.

Use an independent check that follows a different route. Reconstruct totals from components, compare a bound with a point estimate, inspect a distribution behind an average, or solve a changed case with the original definitions. Agreement does not prove every assumption, but disagreement locates a definition, unit or boundary that needs repair.

These sources use technical conventions suited to their settings. This article’s 2×2 arithmetic is original teaching material. It illustrates definitions but does not reproduce a benchmark result or validate a real agent.

Before moving on, record the observation that would distinguish the two most plausible explanations. If no feasible observation exists, narrow the claim instead of pretending the ambiguity has disappeared.

A changed-context transfer check is to compare two source definitions line by line before combining their reported values. State what would change the result and what the example cannot establish. That limit belongs inside the answer.

Contents · Previous chapter · Next chapter

High win rate does not imply low exploitability; a best response is not the average opponent; empirical frequencies are not necessarily the declared policy; exact and approximate gaps differ; and zero exploitability does not guarantee high payoff against every weak opponent.

Match the claim to the evidence. A tidy decimal cannot repair selective observation, and one convenient run cannot establish every product, participant, signal or opponent. Report the result, the conditions that make it interpretable, and the next observation needed for the actual decision. Separate the numerical conclusion from the practical conclusion.

When someone reports 80 per cent wins, ask against whom, under which payoff, from which policy version and whether a best-response test was run. Those questions turn a scoreboard into an evaluation plan.

Before moving on, record the observation that would distinguish the two most plausible explanations. If no feasible observation exists, narrow the claim instead of pretending the ambiguity has disappeared.

A changed-context transfer check is to rewrite a leaderboard entry so it includes opponent set, games, mean payoff and exploitability convention. State what would change the result and what the example cannot establish. That limit belongs inside the answer.

Contents · Previous chapter · Next chapter

Store game definition, payoff units, information structure, policy probabilities or code version, opponent selection, random seeds, number of games, outcome distribution, mean payoff, win rule, best-response method, convergence checks and formula convention.

Begin with a measurement ledger. Name the quantity, its unit, the sample or observation that produced it, the time window and the transformation rule. Mark each input as observed, stipulated, estimated or assumed. This prevents a familiar label from importing a new denominator or a stronger claim. Read the ledger as a sentence before calculating; if the sentence is incomplete, the arithmetic is premature.

An audit finds that the win-rate tournament used one opponent family while exploitability used a different game abstraction. The two numbers should not appear in one ranking until the relationship is justified.

Before moving on, record the observation that would distinguish the two most plausible explanations. If no feasible observation exists, narrow the claim instead of pretending the ambiguity has disappeared.

A changed-context transfer check is to remove the game version from the audit and identify why later reproduction can fail. State what would change the result and what the example cannot establish. That limit belongs inside the answer.

Contents · Previous chapter · Next chapter

Return to the Zero-Sum Games root for payoff, minimax and mixed-strategy foundations. Use the payoff-scenario branch when uncertainty concerns the matrix itself.

Keep the mechanism beside the number. Ask what physical, biological, statistical or strategic relationship connects the observation to the conclusion, and which conditions could break that link. A correct calculation can answer the wrong question when a changed setting alters the process. Draw one arrow for every inferential step, then challenge each arrow with a counterexample.

The transferable habit is to test a system outside the distribution that made it look successful. In security, forecasting, education and operations, observed success and worst-case vulnerability answer different questions.

Before moving on, record the observation that would distinguish the two most plausible explanations. If no feasible observation exists, narrow the claim instead of pretending the ambiguity has disappeared.

A changed-context transfer check is to apply the distinction to a non-game system, naming the observed environment and the stress-test environment separately. State what would change the result and what the example cannot establish. That limit belongs inside the answer.

Contents · Previous chapter · Return to Zero-Sum Games