Chapter 4 · The deep learning era (2006–2016)2016
AlphaGo and deep reinforcement learning
AlphaGo combined deep networks with search and self-play to beat a Go world champion.
Chapter 4 · The deep learning era (2006–2016)2016
AlphaGo combined deep networks with search and self-play to beat a Go world champion.
One copy of the AI plays both sides. Early on, its moves are mostly guesses.
No human games are needed. The AI gets better only from the results of its own games.
Games are great for learning: there are rules, and at the end you know if you won. That final win or loss is the only teacher.
You are X and move first. Choose how many practice games the AI has had. The amber glow shows how much it liked each square on its last move.
Your move (X).
Learned a lot: it avoids many blunders.
Try “AI starts” with 10,000 games: the glow concentrates on the best squares. Untrained, it is spread evenly. This is a simplified teaching version: a lookup table on a 3 by 3 board, not a deep network on a Go board. Counts are for this level only and last until you reload.
What teaches the AI in self-play?
What does the glow on the board show?
What does search add?