untpoker
Contents / Chapter 12
Chapter 12
POKER FROM FIRST PRINCIPLES

Game theory, indifference, and regret

12.1 Nash equilibrium

A Nash equilibrium is a strategy profile in which no player can improve expected value by changing strategy unilaterally while everyone else keeps theirs fixed. In two-player zero-sum poker variants, equilibrium strategy has a particularly strong robustness interpretation: it guarantees the game value against any opponent strategy, even if it may fail to exploit weak opponents maximally.

12.2 Indifference is how mixed strategies are pinned down

Why bluff only some hands rather than all or none? Because equilibrium often makes the opponent indifferent at a boundary. If the defender folds too much, bluffing becomes strictly profitable and the attacker should bluff more. If the defender calls too much, zero-equity bluffs lose and the attacker should bluff less. The stable point is where the marginal bluff earns the same as giving up.

The river formulas from Chapter 7 are a concrete example. With a pot-sized bet, a zero-equity bluff requires 50% folds to break even; a polar betting range that makes a bluff catcher indifferent contains one bluff for every two value hands. These frequencies are dual constraints.

12.3 Exploitability

Exploitability asks how much EV a perfect best response could gain against a strategy compared with equilibrium value. A strategy can win against today's opponent yet be highly exploitable in principle. Conversely, a low-exploitability strategy can leave money on the table against a predictable opponent. This creates the practical tension between robustness and exploitation.

12.4 Regret matching

Imagine repeatedly choosing among actions. For each action k, track cumulative regret: how much better your total payoff would have been if, at each iteration, you had used k instead of the action actually selected, under the relevant counterfactual evaluation. Regret matching assigns future probability in proportion to positive regret. Under appropriate zerosum conditions, average strategies converge toward equilibrium as average regret shrinks.

Regret matching

p(k) proportional to max(R_k, 0)

12.5 Counterfactual regret minimization

Poker is not a single simultaneous move; it is a game tree with hidden information. Counterfactual Regret Minimization (CFR) decomposes regret by information set - states a player cannot distinguish given what they know - and updates action regrets at those local decision nodes. Repeated self-play drives average counterfactual regret downward and can approximate equilibrium strategies in large imperfect-information games.

Historical milestone from the course

Bill Chen's 2015 lecture discusses Cepheus, a heads-up limit Hold'em program whose exploitability was reported below one-thousandth of a big blind, a landmark demonstration that a real poker variant could be solved to extremely small exploitability.

12.6 Why no-limit multiplayer poker is harder

No-limit play adds many bet sizes; deep stacks expand the tree; multiway pots break some of the convenient two-player zero-sum guarantees; and abstraction choices can introduce error. Modern solvers reduce the game through bucketing, action abstraction, subgame solving, or other approximations. The conceptual lesson remains: strategy is a distribution over actions across a tree, not a list of isolated "correct plays."

12.7 What a beginner should take from GTO

  • Use indifference equations to sanity-check bluff and call frequencies.
  • Think in ranges and policies rather than hand-by-hand narratives.
  • Ask what opponent deviation your strategy would be vulnerable to.
  • Use equilibrium as a baseline model, not as a demand to imitate a solver blindly.
  • Exploit only when the evidence for the opponent's deviation is strong enough to justify moving away from a robust baseline.

UNTPOKER · AN INDEPENDENT EDUCATIONAL GUIDE