Hacker News Reader

Comment by afthonos | Hacker News Reader

Important to note that KataGo was double-handicapped. 20 seconds per move maximum; it couldn’t read deep. Against an amateur, it doesn’t matter, but against a historically strong pro it matters a lot.

View on HN

...on a 4x 3090 rig. The game ran 299 moves, giving katago 100 minutes if it exhausted time on each move (which must be the optimal strategy under that time control). Shin used about 205 minutes, over twice as much time and of course had leeway to spend more time on difficult moves.

Based on the youtube video, it looks like katago was only using 16 seconds per move, is that right? https://www.youtube.com/watch?v=-86zF4mTWOY

Is 20 seconds on that hardware really overkill and well into the diminishing-returns curve, as a top-level comment suggested, or is it plausible katago could have played better if given 40 seconds per move?

match details: https://gostonebase.com/blog/shin-jinseo-vs-katago-kishin-ma...

As another question, does it not operate similarly to the top chess engines? The way the neural network systems work is by using the probabilistic matching paired with a Monte Carlo simulation. So you can get to extreme depth very rapidly. Obviously the breadth is going to be limited, but if the neural network side is well tuned (so high probability hits are indeed generally the most challenging moves), then that's not such a problem.

And you can run a huuuuuuuge number of sims in 16 seconds.

I believe the basic idea is similar, but there's an enormous difference in the space of possible moves between chess and Go. Go has a larger board and moves are less restricted. There was a major breakthrough in Go playing programs a decade or two ago when good Monte Carlo methods were developed for it. But still I think the ability to simulate many moves is less powerful in Go.

This is how alphazero worked, but not stockfish iirc.

The tree search in KataGo does not use Monte Carlo simulations at all. It is purely policy/value network based, and works insanely well indeed.

Considering the original AlphaGo ran on a full Google TPU rack, and here KataGo seems to run on a $10k computer, and won with 2 stones handicap, this is showing just how much advance there was in computer go

AlphaGo is a bit difficult to compare because their parallelism efficiency was abysmal. You might as well consider it a regular 8x or so GPU machine.

AlphaGo Master, the much stronger[1] one that beat Ke Jie a year later in 2017, ran on a single 4x TPU (v2?) workstation.

[1] From wikipedia, citing deepmind's paper: 4858 Elo vs 3739 Elo

I've not looked at the specific game in question, but in general, doubling the thinking time leads to a very notable strength improvement regardless of the base level of play. The increase becomes slightly smaller at higher levels but does not really diminish, and it currently looks like the ceiling is still a bit away.

If it is "only" 4x3090 at 16s, you will definitely get a drastic playing strength boost from doubling the thinking time. This is still clearly within the interval of a linear relationship between thinking time and playing strength, i.e. elo ~ log time. The relationship, to my knowledge, becomes less clear only starting at about 10-20x the number of playouts.

Source: Wrote a paper on this. https://ieeexplore.ieee.org/document/10645535

Reply on news.ycombinator.com