Game Theory

In simple terms

In most optimization problems, you control all the variables. In a game, other rational agents also make decisions, and your best choice depends on what they choose — and their best choice depends on yours. Game theory formalises this: given a set of players, their available actions, and payoffs for each combination, what is the rational outcome? The most famous concept — Nash equilibrium — identifies action combinations where no player can unilaterally improve their payoff. Games underpin economics, traffic networks, auctions, protocol design, and increasingly AI.

More detail

Key definitions:

Players: the decision-makers (two or more). Actions / strategies: the choices available to each player. Payoff function: the reward each player receives as a function of all players’ actions. Strategy: pure (deterministic choice) or mixed (probability distribution over actions).

Nash equilibrium: a strategy profile (s_1*, s_2*, ..., s_n*) is a Nash equilibrium if no player can increase their payoff by unilaterally changing their strategy, given all others’ strategies. s_i* = argmax_{s_i} u_i(s_i, s_{-i}*).

Existence theorem (Nash 1950): every finite game has at least one Nash equilibrium (possibly in mixed strategies). This earned John Nash the 1994 Nobel Prize in Economics.

Classic games:

Prisoner’s Dilemma:

         Cooperate  Defect
Cooperate  (3, 3)   (0, 5)
Defect     (5, 0)   (1, 1)

Each player’s dominant strategy is Defect (Defect pays more regardless of what the opponent does), so the Nash equilibrium is (Defect, Defect) with payoff (1,1) — inferior to mutual cooperation (3,3). The dilemma: individual rationality leads to collective irrationality. Models: arms races, environmental agreements, price wars.

Coordination game: two players must choose the same option (e.g., driving on the left vs. right); any consistent choice is a Nash equilibrium. Explains why standards (USB, QWERTY, metric system) are stable once established.

Battle of the Sexes: two Nash equilibria in pure strategies; one in mixed. Coordination problem with conflicting preferences.

Zero-sum games: one player’s gain exactly equals the other’s loss. Chess, poker (simplified). Every finite two-player zero-sum game has a Nash equilibrium solvable by linear programming (minimax theorem, von Neumann 1928).

Dominant strategies: a strategy that is strictly better than all others, regardless of opponents’ choices. If dominant strategies exist for all players, the game’s outcome is determined — the Prisoner’s Dilemma has Defect as a dominant strategy for both.

Mechanism design (reverse game theory): instead of analysing a given game, design the rules (mechanism) to achieve a desired outcome. Examples:

• Vickrey auction (second-price auction): bidders submit sealed bids; winner pays the second-highest bid. Truthful revelation is a dominant strategy — you cannot benefit from lying about your value. Used in Google Ad Auctions, eBay.
• Matching markets: the Gale-Shapley algorithm (deferred acceptance) produces a stable matching (residency assignment, college admissions). Alvin Roth won the 2012 Nobel Prize for designing kidney exchange markets.

Extensive form games: games with sequential moves and information sets — modelled as a game tree. Subgame perfect Nash equilibrium (backwards induction) selects equilibria that are Nash equilibria at every subgame. Chess, negotiation, reputation games.

Repeated games: if a game is played repeatedly, cooperation can be sustained even when single-play equilibrium is defection (the “folk theorem”). Tit-for-tat strategies in repeated Prisoner’s Dilemma outperform always-defect in tournaments.

Game theory in computer science:

• Network routing (Braess’s paradox): adding a road can make everyone worse off in Nash equilibrium. Traffic flows find Nash equilibria, not social optima.
• Multi-agent RL: agents learning in games — GAN training is a two-player zero-sum game (generator vs. discriminator); AlphaGo is a minimax-style player.
• Internet auctions: Google AdWords and real-time bidding (RTB) in ad tech are designed using mechanism design theory.
• BGP routing: strategic interaction between autonomous systems — studied through game-theoretic lens.

Why it matters

Game theory is the mathematical foundation of markets, auctions, negotiations, and multi-agent systems. Mechanism design is how Google’s ad auction generates $170B/year in a incentive-compatible way. Multi-agent AI (AlphaGo, GANs, multi-agent reinforcement learning) uses game-theoretic concepts. Network security (attacker-defender games), protocol design (Byzantine fault tolerance), and distributed systems (consensus with strategic nodes) all use game theory. For AI safety, game theory models multi-agent alignment and cooperative AI.

Real-world examples

• Google AdWords / real-time bidding: generalised second-price auction; Vickrey mechanism design applied at billions of transactions per day.
• Gale-Shapley algorithm: used for NRMP (medical residency matching in the US) and school choice in New York City and Boston.
• GAN training (Ian Goodfellow, 2014): generator and discriminator play a minimax zero-sum game; convergence is a Nash equilibrium.
• AlphaGo / AlphaZero: Monte Carlo Tree Search with minimax — backward induction in extensive-form games.

Common misconceptions

• “Nash equilibrium is the best outcome.” As the Prisoner’s Dilemma shows, Nash equilibria can be collectively suboptimal. The equilibrium is stable (no one wants to deviate unilaterally), not optimal.
• “Game theory assumes purely rational agents.” Classical game theory does. Behavioural game theory (Kahneman, Thaler) models bounded rationality — how real humans deviate from Nash equilibrium predictions.

Learn next

Game theory builds on probability and statistics (mixed strategies) and optimization theory (finding equilibria via LP for zero-sum games). Markov chains model repeated game dynamics. Multi-agent machine learning applies game-theoretic concepts in AI.