Step by Step
S
Sample — run many random simulations
Monte Carlo methods estimate a quantity by running a large number of random simulations rather than calculating an exact answer analytically.
Example: estimating the value of pi by randomly throwing many darts at a square containing an inscribed circle, and using the ratio that land inside the circle.
A
Average — the results converge to the true value
As more random samples are run and averaged together, the estimate converges toward the true underlying value, following the law of large numbers.
Example: averaging the outcomes of 100,000 random dart throws produces a much more accurate estimate of pi than averaging just 100 throws.
⚡
1/sqrt(N) — the specific rate of convergence
The error of a Monte Carlo estimate decreases proportionally to 1 divided by the square root of the number of samples (N). This means quadrupling the number of samples only halves the error — diminishing returns as sample size grows.
Example: going from 100 samples to 400 samples (4x more) roughly halves the estimate's error, rather than reducing it by a full 4x.
Applied Walkthrough
1
AlphaGo needs to evaluate which move is strongest from a given board position, without being able to exhaustively search every possible future game.
2
It uses Monte Carlo Tree Search (MCTS): simulating many random game playouts from each candidate move, and using the average outcome of those simulations to estimate how good that move is.
3
Separately, Bayesian inference problems that are too complex to solve exactly often use MCMC (Markov Chain Monte Carlo) to sample from a complex probability distribution instead.
4
In both cases, the core Monte Carlo principle applies: run many random simulations, average the results, and trust that the average converges toward the true underlying value as more samples are used.
Exam Application
Exams test whether you understand the core Monte Carlo principle (sample, then average, converging to the true value via the law of large numbers) and the specific 1/sqrt(N) convergence rate — meaning that quadrupling samples only halves the error, a frequently tested quantitative relationship. Also expect questions naming specific applications: MCTS (used in AlphaGo) and MCMC (used in Bayesian inference).
⚠ Common Trap
The most common trap is assuming that doubling the number of samples halves the error. Because error scales with 1/sqrt(N), you actually need to QUADRUPLE the number of samples to halve the error — a commonly mistested quantitative relationship.
✓ Quick Self-Check
1. What is the core process behind any Monte Carlo method?
Running many random simulations and averaging the results to estimate a quantity.
Tap to reveal / hide
2. According to the law of large numbers, what happens to the average as more samples are taken?
It converges toward the true underlying value.
Tap to reveal / hide
3. How does Monte Carlo estimate error scale with the number of samples N?
Error decreases proportionally to 1/sqrt(N).
Tap to reveal / hide
4. How many more samples do you need to halve the Monte Carlo estimate's error?
Four times as many samples (quadruple the sample size).
Tap to reveal / hide
5. What famous AI system used Monte Carlo Tree Search (MCTS)?
AlphaGo.
Tap to reveal / hide