Must a mixed strategy use equal probabilities?

No. The example's Column mixture uses one third and two thirds.

Does randomizing guarantee a win?

No. It changes expected outcomes and predictability; individual outcomes can still be unfavorable.

Is alternating choices the same as randomizing?

Not generally. A predictable sequence can reveal the next choice even when its overall frequencies match the desired probabilities.

AI Glossary / Graphs and symbolic AI / Intermediate

Mixed strategy

Definition

A mixed strategy is a probability distribution over a player's pure strategies in a game. Instead of committing to one pure strategy with certainty, the player randomizes according to specified probabilities. In a one-move game, this amounts to choosing among available actions probabilistically.

What it means

A pure strategy specifies what the player does without randomizing. A mixed strategy specifies how likely each pure strategy is to be selected. The probabilities must be nonnegative and sum to one. Assigning probability one to one pure strategy is a special case within this mathematical framework.
Randomization can matter when another player's best response depends on predicting your choice. It is not simply indecision, and it does not mean taking several incompatible actions simultaneously. In a game with multiple stages, a pure strategy can be a complete contingent plan, making a mixed strategy more than a probability distribution over the next visible action.

How it works

Expected payoff is calculated by weighting each possible outcome by the probabilities of the players' choices. In a standard mixed-strategy profile, players randomize independently unless a different correlation mechanism is specified.
To identify an equilibrium mixture, analysts check whether any player can improve expected payoff by changing strategy while others keep theirs fixed. Actions used with positive probability must be equally good responses at equilibrium, although unused actions still need to be checked for profitable deviations. Mixing is a representation of strategy; an arbitrary mixture need not be an equilibrium or an effective choice.

Three practical examples

A game-playing agent can randomize between moves to avoid a predictable pattern in a simultaneous-choice game.
A teaching simulation can compare pure and mixed responses when each deterministic choice has a counteraction.
A strategic-analysis model can examine whether uncertainty about a participant's choice changes another participant's incentives, while keeping assumptions about payoffs explicit.

Think of it this way

Think of choosing a card from a deliberately prepared deck: the mix of cards defines the probabilities before a card is drawn. The analogy is limited because repeated draws without replacement change probabilities, whereas a specified repeated mixed strategy may require fresh independent sampling. A predictable rotation through the cards is also not the same as a private random draw.

Worked example

Consider an invented two-player zero-sum game. Row chooses A or B; Column chooses L or R. Row's payoffs are: A against L gives 2, A against R gives -1, B against L gives -2, and B against R gives 1. Column receives the negative of Row's payoff.
Let Row choose A with probability p and B with probability 1 - p. Against L, Row's expected payoff is 2p - 2(1 - p) = 4p - 2. Against R, it is -p + (1 - p) = 1 - 2p. Setting these equal gives 4p - 2 = 1 - 2p, so p = 1/2. At that mixture, either column action gives Row expected payoff zero.
Let Column choose L with probability q. Row's payoff from pure A is 2q - (1 - q) = 3q - 1. Its payoff from pure B is -2q + (1 - q) = 1 - 3q. They are equal when q = 1/3, again giving zero.
Thus Row mixing equally and Column using probabilities one third for L and two thirds for R form an equilibrium: neither can improve its expected payoff by changing alone. Notice that the equilibrium mixtures are not both uniform.
If Row instead uses p = 0.8, choosing R gives Row 1 - 1.6 = -0.6. Column can exploit that imbalance. The zero equilibrium value concerns expectation, not every individual play: realized Row payoffs remain 2, -1, -2, or 1.

Illustrative case study

In a fictional game-agent project, Alder Arena alternates A and B on successive rounds and calls the policy a 50-50 mixed strategy. The long-run counts are equal, but an opponent that sees the pattern can predict the next move.
The team distinguishes frequencies from the information available before each choice and evaluates a genuinely randomized policy under the intended game assumptions. This invented case illustrates predictability, not a claim that randomization always outperforms a deterministic policy in every environment.

Common misconceptions and mistakes

A mixed strategy need not assign equal probabilities, and it need not use every pure strategy. It is also not automatically a Nash equilibrium. Equilibrium requires checking incentives against the other players' strategies.
Do not confuse uncertainty about what an opponent will do with your own deliberate randomization, although both can be represented probabilistically in analysis. Also, the best response to a known non-equilibrium opponent can differ from an equilibrium policy. The payoff model and information structure determine what conclusions a mixture supports.

Business use

Mixed strategies are useful when analyzing interactions in which predictability affects outcomes. Begin by specifying the available choices, timing, information, and payoffs. Check whether the model is genuinely competitive, cooperative, or a mixture of both; zero-sum reasoning should not be imposed on every business interaction.
Treat computed probabilities as conditional on those assumptions. Evaluate sensitivity when payoffs are estimates and distinguish theoretical equilibrium from observed behavior. For deployed decision systems, document how random choices are generated and assessed. A probability table is not evidence that participants actually behave according to it or that the underlying game captures all relevant incentives.

Check your understanding

Why is Row's 50-50 mixture not enough by itself to claim that any pair of strategies is an equilibrium?

Answer: Equilibrium also requires Row to have no profitable deviation against Column's strategy. In the example, Column must use the corresponding one-third/two-thirds mixture to make Row's two actions equally good.

Related terms

Pure strategy
Zero-sum games
Minimax
expected payoff
Nash equilibrium
best response
zero-sum game
randomized policy
correlated equilibrium
behavioral strategy

Learn next

Recalculate Row's payoff against each column choice for p = 0.25 and p = 0.75. Then verify the equilibrium using all four outcome probabilities. Keep the probability distribution, realized action, and expected payoff distinct throughout the calculation.

Probability distribution

Sources

Constantinos Daskalakis and Gabriele Farina, original MIT Topics in Multiagent Learning notes, normal-form games and mixed strategies: https://www.mit.edu/~6.7980/nfgs_nash.html

Previous: Mixed sampling Next: Mixing gradients