Are playout moves always random?

No. Policies can use heuristics or learned preferences.

Is one playout the whole search?

No. Tree search typically combines many evaluations with selection and statistical updates.

Does a draw always count as half a win?

No. That is the explicit reward convention in the worked example; another task can define rewards differently.

AI Glossary / Graphs and symbolic AI / Advanced

Playout

Definition

A playout is a simulated continuation from a game state to an outcome, with moves selected according to a specified policy. In Monte Carlo tree search, it is often called a rollout. Repeated playouts provide outcome samples that help estimate the value of a position or action. A simulated win is evidence about that continuation, not proof that the starting position wins against every opponent.

What it means

Instead of examining every possible future, an agent follows selected moves through a simulated game. It records the result and repeats the process. A playout policy determines which legal moves are selected along the way; it may be random, use simple rules, or incorporate learned guidance.
Results therefore depend on the policy used for both sides. Many wins against weak simulated responses can overstate how promising a move is against a strong opponent. The rules simulator, stopping condition, and reward convention also affect what the samples mean.

How it works

In a basic Monte Carlo tree search cycle, a tree policy selects a path through explored positions. After expansion, a playout can continue from a selected position to a terminal state. The result updates statistics along the searched path, helping later iterations decide where to explore.
A full playout reaches the game's end. Some systems instead stop at a depth or computation limit and use an estimated value; such a truncated rollout includes evaluator error rather than a directly observed terminal result. Systems using value estimates need not perform full-game playouts at every evaluation.
The original AlphaGo combined learned networks, tree search, and fast rollout policies. This illustrates that playouts need not use uniformly random moves.

Three practical examples

A board-game agent samples legal continuations after a candidate move and averages their rewards. The policy governing the opponent's simulated moves influences whether the estimate captures tactical dangers.
A game designer compares two rule variants by simulating games with defined player policies. Different outcome frequencies may reflect the policies' strengths or preferences, so they are not automatically evidence of balance among expert human players.
A training environment evaluates an agent by running complete simulated episodes. These trajectories can resemble playouts, although the broader term episode is common outside game-tree search.

Think of it this way

Think of rehearsing possible conversations before an important meeting. Rehearsals can reveal useful consequences, but imagining only agreeable responses gives a misleading picture. A playout is similarly informative only in relation to how its continuation is chosen. Running the same unrealistic rehearsal more often does not correct its assumptions.

Worked example

Consider a fictional two-player game scored from our agent's perspective: a win earns 1, a draw 0.5, and a loss 0. Two candidate moves each receive twenty illustrative playout results under the same simulation policy.
Move A produces twelve wins, four draws, and four losses. Its total reward is 12 + 4 × 0.5 = 14, so its mean is 14/20 = 0.70.
Move B produces thirteen wins, no draws, and seven losses. Its mean is 13/20 = 0.65. A has the higher average reward despite fewer wins, because draws contribute half a point. If the objective counted only wins, B would instead have the higher observed win rate. The reward definition changes the comparison.
Now examine a separate simplified position. After move C, the opponent has ten legal replies: nine lead to our win and one forces our loss. Uniformly random opponent replies give C an expected reward of 0.9. Move D guarantees a draw, worth 0.5, against every reply.
Random-response playout averages favor C. An opponent who always selects the refutation makes C worth zero, while D remains worth 0.5. The random-policy expectation and the worst-case value answer different questions.
With twenty independent uniform samples of the opponent's reply after C, the probability of never sampling the refutation is 0.9 raised to the twentieth power, approximately 12.2%. More sampling reduces that particular miss probability, but merely averaging uniformly random replies still estimates the random-response value. Discovering a refutation and using it appropriately in adversarial search are additional steps.

Illustrative case study

Fictional game studio Willow Table builds an automated opponent using simulated continuations. A candidate strategy wins frequently against its initial rollout policy, but human testers repeatedly use a defensive reply that the simulator rarely chooses.
The team adds tests around those positions, reviews the opponent policy, and compares performance under a fixed computation budget. It records completed playouts, selected moves, and results against separate test opponents. It also checks that a reward recorded for one player is not accidentally interpreted as the other player's reward. This fictional case illustrates validation of simulation assumptions rather than a claim that a particular policy change always improves play.

Common misconceptions and mistakes

A large playout count does not guarantee optimal play. The search method, simulation policy, and available computation matter.
A sample average is not automatically a calibrated probability of beating the real opponent.

Repeated or dependent samples do not provide the same information as independent draws under an unchanged policy.
A truncated simulation should not be reported as a completed game if its result comes from an evaluator.

Business use

For organizations using simulation-based decision systems, the useful lesson is to make continuation assumptions explicit. Document how actions are selected, how outcomes are scored, and when simulations stop. Compare methods under the same relevant resource budget, since a more elaborate playout policy may produce fewer samples in the available time. Test failures that average results can hide, and distinguish performance in a simulator from outcomes in the intended environment.

Check your understanding

Why does A outscore B in the example?

Answer: Its four draws raise its total reward to fourteen versus thirteen.

Which move is better against the refuting opponent, C or D?

Answer: D, because it guarantees a draw while C loses.

Does more uniform sampling change C's random-response expectation of 0.9?

Answer: No; it improves estimation of that expectation rather than changing the policy being evaluated.

Related terms

Monte Carlo tree search
Terminal states
Minimax
Reward
Rollout
simulation policy
tree policy
terminal state
value estimate
sample average

Learn next

Recalculate A and B when a draw earns 0.25. Then explain why a policy that finds tactical refutations might change action rankings even if it completes fewer playouts per second.

Sources

Silver and colleagues, Mastering the game of Go with deep neural networks and tree search: https://storage.googleapis.com/deepmind-media/alphago/AlphaGoNaturePaper.pdf AIMA glossary, playout: https://github.com/aimacode/aima-glossary/blob/master/glossary.md

Previous: Platykurtic Next: Playout policy