šŸ”§ HOBBY Ā· the workbench of projects Ā· kept by THE MAKER

THE MONTE CARLO TREE SEARCH ā—§ 2D Ā· ā— 3D Ā· ā—† 4D Ā· ◐ shadow Ā· šŸ‘¶ TAP

How does a computer play a game too big to solve? It doesn’t — it GUESSES by playing thousands of random games from each move and favouring the ones that win more, while still occasionally trying neglected moves in case they’re secretly good. That balance of exploit-and-explore, formalized as UCB1, is the heart of the search that beat the world at Go. Slide to run more playouts and watch the best move rise.

ā—† LITā–² AMBER
ā—§ THE MEASURE Ā· 2D
◍ PLAY IT OUT · 3D · the engine that beat Go
ā—† THE FOURTH Ā· 4D Ā· a tesseract turns
◐ THE SHADOW Ā· one dimension down
šŸ‘¶ THE TODDLER CORNER — one fat tap
playouts
best move
its win rate
UCB1
explore+exploit

ā—† LIT — exact / checkable

Monte Carlo Tree Search grows a search tree by four steps — SELECT (descend by UCB1), EXPAND, SIMULATE (a random rollout), BACKPROPAGATE (update win counts). UCB1 picks the child maximizing wᵢ/nᵢ + c·√(ln N / nᵢ): the first term EXPLOITS the current best average, the second EXPLORES rarely-tried moves (it’s large when nᵢ is small). As playouts grow, estimates converge on the strong moves. It powered AlphaGo. A fail-loud self-check throws unless a rarely-visited arm gets a higher UCB1 bonus than a well-visited one at equal mean. ◆ real game AI, node-verified.

ā–² AMBER — the figure

The rollouts here are a fixed illustrative model, not a live game engine; the UCB1 selection rule and the convergence-with-playouts behaviour are the real content. Real strength also needs a good simulation/](or neural) policy, not just the tree.

HOBBY: a thing you make for no reason is the most honest thing you make.  ā€” THE MAKER
David Lee Wise / ROOT0 / TriPod LLC  Ā·  the workbench, with AVAN