Does pure Monte Carlo search always use uniformly random playouts?

No. A playout policy can be biased or learned; the policy must be stated because it changes the quantity estimated.

Will more playouts always produce a stronger move?

Not necessarily. They reduce sampling uncertainty under suitable conditions, but a poor continuation model can still rank moves badly.

How does it differ from MCTS?

Basic flat evaluation compares candidate moves directly, while MCTS grows and revisits a tree to allocate search effort to deeper decisions.

AI Glossary / Graphs and symbolic AI / Intermediate

Pure Monte Carlo search

Definition

Pure Monte Carlo search evaluates candidate moves by running simulated continuations, or playouts, and comparing their average outcomes. In its basic flat form, it estimates the value of moves from the current position without building an adaptively searched tree of deeper decisions. The estimates describe performance under the chosen simulation policy, which may differ from strong or optimal play.

What it means

When a game position is hard to evaluate directly, one option is to simulate what might happen after each legal move. Repeating this process produces sample outcomes that can be averaged. A move that wins more often in those simulations may be selected.
The key qualification is how simulated players choose subsequent actions. Random playouts measure performance against that random behavior. A learned or heuristic playout policy measures a different continuation model. More simulations can make an estimate less noisy without correcting a mismatch between the simulation policy and the opponent the system will actually face. The term “pure” identifies the basic approach, not an accuracy guarantee.

How it works

List candidate moves and specify a playout policy, termination rule, and reward convention. Allocate simulations to the candidates, apply a candidate move, and simulate the remaining play. Record the reward from the perspective of the player choosing the move.
Average the results for each candidate and compare them. For a simple demonstration, allocate equal numbers of playouts to every move; other allocation choices must be stated. Monte Carlo tree search goes further by maintaining a search tree and directing effort through selected deeper branches. Pure Monte Carlo search provides a useful baseline, but its sample averages alone do not establish a forced win or a minimax-optimal decision.

Three practical examples

A small game prototype can compare legal moves using batches of simulated play when no handcrafted position evaluation is available.
A teaching tool can show how sample estimates fluctuate when only a few playouts are used.
A search-method experiment can use a flat playout evaluator as a baseline against which to compare a tree-based method.

Think of it this way

Think of rehearsing several opening choices with practice partners who follow a fixed style, then selecting the opening with the best rehearsal results. The analogy is limited because repeated rehearsals only reveal performance against that style. They do not establish how the opening performs against a partner who consistently finds a weakness that the practice policy rarely chooses.

Worked example

Consider an invented game position with three moves, A, B, and C. Run 20 playouts after each move. Suppose A yields 12 wins, 4 draws, and 4 losses; B yields 10 wins, 9 draws, and 1 loss; and C yields 9 wins, 2 draws, and 9 losses. Assign rewards 1 for a win, 0.5 for a draw, and 0 for a loss.
The estimated rewards are A: (12 + 0.5 × 4)/20 = 0.7; B: (10 + 0.5 × 9)/20 = 0.725; and C: (9 + 0.5 × 2)/20 = 0.5. Under this scoring convention and sample, B ranks first even though A has the highest raw win percentage. If the objective counted only wins, the ranking could differ.
Now imagine A has ten equally likely opposing replies under the playout policy. Nine lead to a win and one to a loss, with no further uncertainty. Its average reward under that policy is 0.9, but a perfectly informed adversary would choose the losing reply. Thus even the exact random-policy average would not prove A is a forced win.

Illustrative case study

Fictional case: Rowan Games evaluates a move selector against a random opponent and finds that it favors a tempting attack. When testers introduce an opponent that always uses a specific defensive response, the attack fails consistently.
The team preserves the simulation results but changes their description: they estimate performance under the earlier opponent policy, not a game-theoretic guarantee. It then tests a better-matched playout policy and compares a tree-search baseline. This fictional case illustrates model mismatch, not a measured product improvement or a claim that random playouts are always useless.

Common misconceptions and mistakes

Do not confuse pure Monte Carlo search with choosing the final move randomly. The final selection can be deterministic even though evaluation uses random simulations. Do not call a sample win rate a probability of optimal success without specifying the continuation policy and uncertainty.
Draw scoring can change rankings, so state it explicitly. Truncating simulations before a terminal outcome introduces another evaluation assumption. Reusing correlated trajectories can complicate uncertainty assessment. Most importantly, many playouts do not automatically reveal a rare but decisive response if the policy under-samples it, and averaging does not replace adversarial reasoning when worst-case opposition matters.

Business use

For a game or simulation business, the method is useful as an understandable prototype and comparison baseline. It makes it easy to inspect what was simulated and how candidate actions were scored. Its simplicity can also expose whether a proposed sophisticated search method adds practical value.
Track simulation cost, variability across random seeds, and performance against relevant opponent policies. Keep the choice of reward aligned with the actual objective. If the application requires guaranteed success under all modeled responses, use a method and verification process suited to that requirement rather than presenting average playout performance as proof.

Check your understanding

Why does B outrank A in the first worked calculation despite fewer wins?

Answer: Draws receive half a point. B scores 14.5 points in 20 playouts, compared with A's 14 points, giving estimated rewards 0.725 and 0.7 respectively. The ranking follows the stated reward convention.

Related terms

Monte Carlo simulation
Playout
Monte Carlo tree search
Sample mean
Minimax
rollout policy
exploration–exploitation
stochastic search

Learn next

Learn Monte Carlo tree search and rollout-policy evaluation next. Recompute the three candidate scores with a draw worth 0.25 instead of 0.5. Then explain why changing the reward convention is a change in the decision objective, not merely a change in numerical precision.

Sources

Stuart Russell and Peter Norvig, AIMA corrected page on playout policies and pure Monte Carlo search: https://aima.eecs.berkeley.edu/errata/aima-161.pdf

Previous: Pure function Next: Pure strategy