AI best choice

1 Like

So, this AI analysis shows the best moves next, one is 1.1, one is 0.8, but the 0 is highlighted. I thought 1.1 is the biggest winrate increase move, right? But then why 0 position is highlighted?

1 Like

I guess blue is move with the most playouts, not winrate.
AI do more playouts on better moves, so their greatness has big statistical proof.
But score on moves with few playouts changes much sometimes because not enough playouts, they are possibly better, but currently AI is less sure of them, more playouts needed.

2 Likes

What is playout? I keep seeing this term, but never really understand it.

2 Likes

Modern Go AI uses a combination of deep learning and Monte Carlo tree search. The tree search is a method of evaluating a move with random simulations. The algorithm selects a possible move, and then simulates many games to completion from that new position, using random moves for both sides. After many games, if the win rate is higher for the side it is evaluating the move candidate for, that move might be good. Each of these simulated games is called a “playout”. This evaluation will be much better if you simulate many more games of course, so more playouts means more accurate evaluation. But also more costly to compute.

I’m sure I am missing some nuance as I am not an expert, but this is the basic idea. This is a good book to go deeper (pun intended) Chapter 4. Playing games with tree search · Deep Learning and the Game of Go

9 Likes

There’s also this in the OGS doco/wiki: Understanding the AI review · online-go/online-go.com Wiki · GitHub. Bots aren’t really my thing so a link is about all I’m good for.

3 Likes

A playout is a search kind of. So when the computer plays - out it searches the board position or game and yes michiakig has a more detailed explanation than me :slight_smile:

In your board position 1.1 is probably the strongest move suggested by Ai during it’s search. While the 0 is it’s policy or something with just lower readouts/playouts/visits I believe.

In this case 1.1 would be your best but 0 would be the policy’s best choice.

Got it. thanks.

if i cheated, i would pick the one i understood. oh wait, in that case, i just need to keep cheating. lol

it is “strongest”, but if it has like 20 experiments, then you shouldn’t trust result of too few number of experiments. But OGS don’t writes how many playouts each move has, so blue is the most trustful move.

2 Likes

Really? Well that’s good to know now :slight_smile: