How did AlphaGo use supervised learning and reinforcement learning?

How did AlphaGo use supervised learning and reinforcement learning?

The original AlphaGo combined learning from human games with reinforcement learning through self-play. It also used a search procedure to examine possible future moves. The important idea was the cooperation between learned guidance and structured search.

Learn useful moves from human play

A policy network was trained on expert moves to predict which move a strong player would choose in a position. This supervised stage supplied an effective starting point.

Reinforcement learning then improved a policy through games of self-play, using winning as the goal. A value network learned to estimate the outcome of positions, providing another kind of guidance. These components are described in the 2016 AlphaGo paper.

Use networks to guide a search

Policy information helped direct attention toward promising moves, while value estimates helped evaluate positions without exhaustively playing out every possible future.

The system combined this guidance with Monte Carlo tree search. Go's branching possibilities make exhaustive search impractical, so better guidance can make limited search computation more useful.

Imagine evaluating ten plausible moves before spending effort on hundreds of unlikely ones: learned estimates help decide where that effort belongs.

Distinguish AlphaGo from AlphaGo Zero

AlphaGo Zero, introduced later, learned through self-play without using human game examples.

These systems illustrate different training designs. Neither should be described as a language model, and their success in a precisely defined board game does not automatically establish reliable reasoning in open-ended conversations.