Neural (Experimental) — A Self-Learning Quoridor Engine
Neural (Experimental) is the odd one out. Every other opponent on the roster is the same search at a different budget; this one is a different program entirely, built around a neural network that learns from games rather than a procedure that was written down once. It is also, right now, the weakest bot here — not the hardest. It is the right opponent if you have just finished your first game and want something you can actually beat, or if you are curious what a learning engine looks like before it is any good.
That combination — genuinely different, genuinely weak — makes it the best place on this site to look at a distinction that matters far more for your own play than it does for the machine's: the difference between judging a position and calculating one.
Two ways to know a move is good
Ask the six-tier ladder whether a move is good and what comes back is a sampled outcome (how the engine scores a move). Ask this one and what comes back is an opinion. Its network looks at the board once and returns two things: a rating for every legal move, and a single number estimating how the position turns out for whoever is to move. The search built on top of that spends its time following the moves the network rated highly and revising those estimates as it explores — but the values it revises them with come from the same network that produced them in the first place.
So the ladder calculates. This bot judges, then checks its judgement against more judgement. Everything below follows from that.
What its budget actually is
Its per-move allowance is bounded two ways at once: a target number of search steps, and a wall clock, whichever runs out first. In a browser the clock always wins. The step target is set generously and is simply not reached, because every position this engine looks at costs a full network evaluation instead of a cheap simulated game — the ladder's upper tiers get through something like ten times as many positions per second of thinking.
The consequence is worth knowing before you sit down. It thinks for close to eight seconds a move, every move, from the first to the last. That makes the weakest bot on the site comfortably the slowest one: roughly twice the per-move thinking time of the hardest tier on the ladder, which is done in under four seconds. A quick game against the easy bot is not a quick game. Its allowance is cut further by whatever is left on the game clock, as every bot's is (what a fast clock does to a search) — but here that is the ordinary case rather than a caveat, since the clock is what stops it anyway.
Two practical things. The model file downloads the first time you ask it for a move, so the opening move of your first game is slower still; if that download fails it tells you so rather than quietly handing you a move from a different engine. And it plays two-player games only — it is not offered in four-player games, which run on a different engine altogether (see the FAQ).
It also has no opening book. Where the rest of the roster plays its first few moves out of a fixed table (the opening is a lookup), this one is on its own from move one and produces its opening moves exactly the way it produces every other move. A settled race with no walls left is answered directly rather than searched (how the engine works); the arithmetic behind those positions belongs to Endgames.
What it finds, and what it misses
It finds moves that look right. Its move ratings were trained partly on an archive of real games played on this site, and against fixed reference opponents that shows: it beats a random pawn-walker every game, and it beats an opponent that does nothing but walk its own shortest path every game — which means it does understand that walls exist and roughly where they belong. Against the gentlest tier of the actual ladder it takes about one game in six. That is the honest shape of its ability: real, and small.
What it misses is anything whose value only shows up several moves later. Because its verdict on a position is a single learned number rather than a played-out result, a position it judges to be fine can already be lost. Concretely, it fails the trap-and-cage test positions this project keeps for exactly that purpose, and it fails them across the board rather than occasionally. Worth knowing that the strongest bot on the ladder misses on the same suite too — far less badly, and Nemesis's page carries the measurement and what it means. But wall structures that close around a pawn over two or three moves, the material of Traps and Cages, are this bot's clearest blind spot by a distance.
Its weakness is also real, not manufactured. The gentlest tiers on the ladder are deliberately made to blunder, which Scout's page covers; nothing of the kind is applied here. This engine is simply not good yet, and every mistake it makes is one it meant.
The lesson: judgement is cheap, calculation is decisive
Your own play mixes the same two modes. Most of your moves are judgement — this looks like the right side of the board, that wall looks strong, I feel ahead. A few are calculation: counting both routes and comparing the two numbers, as in Tempo and the Race. Judgement is not the enemy. It is how you get through a game in finite time, and strong players judge far more often than they count. The failure mode is not knowing which mode you are in, because a move you judged and a move you counted feel identical from the inside, and only one of them is reliable.
This bot is that failure running with almost nothing behind it, which is what makes it worth playing even though it is weak. Watch how many of its moves look completely sensible right up until the moment you count them, then ask how many of yours are the same. The practical version: judge freely, but calculate at the moments when the judgement is about to be spent — before you place a wall, and whenever the position feels close.
How to beat it
Race, and count while you do it. Its weakest reading is of positions whose value arrives late, so give it a position whose value arrives steadily. Walk forward, keep both route counts in your head, and make it prove it can stop you. It usually cannot.
Do not be impressed by purposeful-looking walls. Its move ratings are good at producing plausible walls, and plausible is not the same as effective. Recount your own route after each wall it places rather than assuming it did what it appeared to do. Often you will find it added nothing at all, and you have just been handed a free tempo by an opponent who spent a turn.
Build the thing it cannot see. If you want to end the game rather than merely win the race, aim two walls at the same detour and close a pocket around its pawn. That is the structure it evaluates worst, and it will often walk into the second wall having taken no precaution against it.
Expect it to repeat itself, mostly. Nothing random is added when it picks a move — no noise, no coin-flip between two moves it rates equally — so from the same position it will usually play the same thing. It is not a guarantee: the search is stopped by the clock rather than after a fixed count, so a faster or less loaded machine gets further into the tree and can land somewhere else. Treat a line that beat it as likely to beat it again, which is entertaining once and teaches you nothing the second time. Beating a mostly fixed opponent repeatedly is not evidence that you improved.
Generations, and why "newer" does not mean "stronger"
The bot improves in generations, not continuously. Each candidate is trained away from the site on a mix of self-play games and an archive of real games played here, and it only replaces the live one if it clears a bar. Your game is not retraining the opponent mid-match, and the bot you play today will be the same bot tomorrow.
The bar is the interesting part, and it is stricter than it sounds. Two candidate generations trained after the one you are playing beat it head-to-head and were still not shipped, because when each was measured against the fixed outside opponents it came out weaker than the version already live. One of those two was the best predictor of human moves the project has produced and, on that external measure, among its weakest players. It got measurably weaker the longer it was allowed to think: given a small search it beat a reference opponent every game, and given the full production budget it lost to the same opponent every game. Its sense of who was winning was miscalibrated, so more searching did not correct the error, it compounded it.
That is this page's own point playing out in the machine's life rather than yours. Winning a match against one particular opponent is a judgement about strength; holding up against a fixed external standard is a calculation of it, and the two disagree far more often than anyone expects.
The direction of travel is real but slow. The current generation has started taking the occasional game off the gentlest tier on the ladder, having previously taken none at all, and still loses the large majority of them. That is what genuine progress looks like from the bottom, and it is why the bot is labelled experimental rather than quietly rated as though it were finished.
Who should play it
Play it if you are new and want a game you can win, or if you want to watch a learning system be honestly bad at something in public. Do not use it to measure yourself: beating it is a narrow skill, and a slow game against the weakest engine here tells you nothing about how you handle an opponent that punishes a loose wall instead of failing to notice it.
When you want that opponent, move up to Scout and take the ladder from there. Current strengths for every bot are on the roster page, and all of them are free to play without an account — paid plans add coaching and analysis on top of the game rather than gating it.