In Go, a board position can offer hundreds of legal moves. After each move there are hundreds more possibilities. DeepMind's AlphaGo used learned estimates to decide which branches deserved attention and which positions looked promising.
The 2016 paper describes a policy network that helped pick moves worth considering and a value network that estimated which side was likely to win from a position. Expert human moves provided an early training signal. The researchers also used self-play, in which a program plays games against versions of itself, to improve play and learn to evaluate positions.
Learning helped the search spend its time
The system did not have to choose a move from the network's first suggestion. It could simulate possible continuations and use what it learned to guide that exploration. The paper reports a 5–0 victory over European champion Fan Hui on a full-size board, before AlphaGo's better-known match against Lee Sedol.
The combination mattered. Training gave AlphaGo a way to judge possibilities. Search let it check some of those possibilities in more detail before committing to a move.