← All games

Glossary

The words this site uses, defined once each

Every term below is one of the cards the front page hands out a day at a time. They are collected here because a word is more useful when it can be looked up than when it happens to come round.

The vocabulary is of two kinds and they are worth telling apart: some of it belongs to machines that play — how they search, what they store, where they stop — and some to the human side being measured against them. Both are here, because the contest needs both halves to make sense.

ABCEIMOPQSTWZ
A mill, three in line
Three of your own pieces in a row along a line of the board. A mill removes one enemy piece. The name is about repetition: the row can be opened and closed again, and each new closing takes another piece, so the mill grinds for as long as the opponent has something to pay with.
Alpha-beta pruning
A way to stop examining a branch already known to be worse than one you have: if the opponent has a reply there stronger than you can accept, the rest does not matter. The answer is unchanged; the work is not. With good move ordering it roughly doubles reachable depth.
Knuth & Moore, Artificial Intelligence, 1975
Amodal completion
Seeing a whole shape where part of it is hidden behind something else. A sheet of paper behind a mug is seen as one sheet, not two separate strips. The completed part carries no colour and no brightness, hence amodal: it belongs to the object and not to the image on the retina.
Michotte, Thines & Crabbe, Les complements amodaux des structures perceptives, 1964
Blunder, a lapse of attention
A move that loses on the spot for a reason the player can explain a second later. A blunder differs from a mistake: a mistake needs analysis, a blunder needed only attention. Machines do not blunder in this sense — they have no fatigue and no moment when the eye slid past the relevant line.
Change blindness
Failing to notice a large change in a scene when the moment of change is masked: by a blank frame, a blink, an eye movement or a door carried past. The change is fully visible and simply does not reach attention. It follows that no detailed picture of the scene is kept in the head.
Rensink, O'Regan & Clark, Psychological Science, 1997
Endgame tablebase
A table holding, for every position with few pieces, the exact result and the number of moves to it. Built by retrograde analysis: backwards from the mates. Here a program neither evaluates nor searches — it reads off the answer. The one stretch of the game that is closed for good.
Thompson, ICCA Journal, 1986
Evaluation function
The number a program substitutes for what it did not have time to compute. Material, mobility, king safety, structure — all collapsed into one value at a leaf of the tree. Human beliefs about the game live here, mistakes included: the search is exact, the evaluation never is.
Iterative deepening
Search two plies, then three, then four, until the clock runs out. It looks wasteful, but the last layer of the tree already outweighs all the earlier ones together. In return an answer exists at every moment, and last pass's move ordering speeds up the next.
Monte Carlo tree search
Instead of full enumeration, the program plays a game out to the end many times with random or near-random moves, and grows the tree where the results look promising. No evaluation function is needed; the statistics of the playouts are the evaluation. It moved Go forward at last and became the core of AlphaGo.
Coulom, Computers and Games, 2006
Opening book
A prepared list of opening moves, usually assembled by people from human games. While the position is in the book the program does not search at all; it recalls. This is the one place where borrowed experience sits openly inside a machine's play, which is why AlphaZero dropped it.
Ply, the unit of depth
One move by one side. What chess notation counts as a single move is two plies to a program. Search depth is measured in plies, and everything hangs on it: each extra ply multiplies the work by roughly the number of moves available in a position.
Poisoned pawn
A pawn that can be taken but should not be: the bill arrives several moves later. A classic depth test — a program with a short horizon grabs the material and dies. A human usually declines because he has seen it before, not because he calculated it.
Pop-out search
The property of a target that differs by one simple feature: it is found in about the same time however many distractors there are. A red circle among green ones pops out; a red circle among red squares and green circles has to be searched for, and the time grows with every extra item.
Treisman & Gelade, Cognitive Psychology, 1980
Quiescence search
The answer to the horizon effect. On reaching its depth limit a program does not stop at once but plays out captures and checks until the position goes quiet. Only quiet positions may be evaluated: in the middle of an exchange any evaluation is off by a whole piece.
Self-play, games against yourself
Learning from games against yourself. The opponent is always exactly your own strength, so difficulty rises with skill and no human games are needed. The method predates neural networks: Samuel's program learned this way in the fifties, TD-Gammon in the nineties.
Subitizing, counting without counting
Knowing how many there are without counting. It works up to three or four objects: the answer arrives at once, exactly, at almost no cost. Past that a different mechanism takes over — approximate estimation whose error grows with the number. Response times show exactly where the border is.
Kaufman et al., American Journal of Psychology, 1949
Tempo, a step ahead
The unit of being a step ahead: who is forced to open up first. Tempo is not material, it does not show in the piece count, and it still decides games. Dots and Boxes controls it openly — giving away two boxes buys the right to make the opponent open the next chain.
The Elo rating
A scale on which four hundred points of difference mean roughly a tenfold edge in expected score. The number has no absolute meaning: it measures relations inside one pool of players. Which is why an engine's rating and a human's cannot simply be subtracted.
Elo, The Rating of Chessplayers, Past and Present, 1978
The flying king
A man that has reached the far row. In Russian draughts it moves and captures along a diagonal for any distance in any direction, which is why it is called a flying king. One king is often worth three men, and most of the game is built around who gets there first.
The horizon effect
A program searches to a fixed depth, and whatever happens beyond it does not exist for the program. A coming loss can be pushed past the horizon with pointless checks, and the position will look fine. The trouble is not a wrong evaluation but nothing left to evaluate.
The minimax rule
The rule for choosing a move: assume the opponent will reply as well as they can, then take the line whose worst outcome is still acceptable. Neither optimism nor caution — an assumption of a flawless opponent. Against a person who errs it is often too modest.
von Neumann, Mathematische Annalen, 1928
The Stroop effect
Naming the ink colour is harder when the word names a different colour. The delay is around a hundred milliseconds and it is stubborn: in a literate adult, reading starts by itself and cannot be switched off on request. The automatism that saves you time all your life works against you here.
Stroop, Journal of Experimental Psychology, 1935
Transposition table
Different move orders lead to the same position. A program keeps positions it has already searched in a large table and, on meeting a familiar one, takes the stored answer instead of searching again. The tree becomes a graph, and the work drops noticeably.
Weber fraction
The ratio of the just-noticeable difference to the quantity itself. It stays constant where absolute precision does not: telling ten dots from eleven is about as hard as telling a hundred from a hundred and ten. Hence the rule that the error of an estimate grows in proportion to the number.
Fechner, Elemente der Psychophysik, 1860
Working memory
Not a warehouse but a workbench: what you are operating on right now. Storage and processing draw on one resource, so holding while transforming costs more than holding alone. Span tasks and updating tasks measure different sides of it and return different numbers for the same person.
Baddeley & Hitch, Psychology of Learning and Motivation, 1974
Zugzwang, forced to move
A position where every available move makes things worse and passing is not allowed. What loses is not the position but the obligation to move. Nim states the case exactly: from a position whose exclusive-or is zero, every move leads to a lost one, and there is no way around it.
Where to go from here
To cite this page: “Glossary”, You vs Computer, https://youvscomputer.com/games/glossary.html