AlphaGo Zero: Learning from scratch

Re:AlphaGo Zero: Learning from scratch by espadrine on Wednesday October 18, 2017 @05:16PM (#15503202)

Random rollouts are what the MC in MCTS stands for.

Without that, it is simply a tree search.

Excerpt from the paper:

> [AlphaGo Zero] uses a simpler tree search that relies upon this single neural network to evaluate positions and sample moves, without performing any Monte-Carlo rollouts.

[ Parent | on HN ]

Re:AlphaGo Zero: Learning from scratch by xoroshiro on Wednesday October 18, 2017 @08:30PM (#15504510)

Does this mean it learns what to search? I wonder why they thought it was a good idea. I thought the whole point of MC was that pruning algorithms like the ones in chess wouldn't work for a larger search space.

[ Parent | on HN ]

Re:AlphaGo Zero: Learning from scratch by letlambda on Thursday October 19, 2017 @01:31AM (#15505665)

The policy network is a function from board states to a scoring of moves. The policy network with the greedy heuristic, ie pick the highest rated move with no explicit look ahead method, plays at a high amateur level.

This was... unexpectedly good.

It effectively reduces the branching factor of Go from the number of moves available, to the number of moves actually worth considering.

[ Parent | on HN ]

Re:AlphaGo Zero: Learning from scratch by shmageggy on Wednesday October 18, 2017 @09:45PM (#15504835)

That's what the policy is. Given a board state, the policy gives you a distribution over all available moves.

[ Parent | on HN ]