[Houshalter]

Hey, they provided the human players with a power outlet and let them draw as much power as they wanted. The competition was totally fair.

But seriously, it's possible that AlphaGo is already much more energy efficient than a human player. The main reason it uses tons of energy, is the tree search part of the algorithm. Where it runs hundreds of thousands of simulated games to further analyze every move. This improves it's skill, but only by a little bit. IIRC, the version without tree search beat the full version 25% of the time. Which would still give it a higher elo than Sedol, which only beat it 20% of the time (and AlphaGo has improved since those games.)

Google is also using custom TPUs, which are claimed to be something like an order of magnitude or more energy efficient than GPUs. And computing technology is only getting more energy efficient with time. In principle, transistors moving around a few electrons are vastly more energy efficient than the very wasteful chemical reactions used in brain. We also know how to "sparsify" nets and remove tons unnecessary connections that could reduce computations a lot. But there's generally no point in doing that because it's not faster on normal hardware.

[comboy]

> IIRC, the version without tree search beat the full version 25% of the time.

That would be amazing but it seems hard to believe. Any references?

I found this (which is also impressive):

    AlphaGo team then tested the performance of the policy 
    networks. At each move, they chose the actions that were 
    predicted by the policy networks to give the highest 
    likelihood of a win. Using this strategy, each move took 
    only 3 ms to compute. They tested their best-performing 
    policy network against Pachi, the strongest open-source 
    Go program, and which relies on 100,000 simulations of 
    MCTS at each turn. AlphaGo's policy network won 85% of 
    the games against Pachi! 

1. https://www.tastehit.com/blog/google-deepmind-alphago-how-it...

2. https://gogameguru.com/i/2016/03/deepmind-mastering-go.pdf

[Houshalter]

I believe I was remembering this from wikipedia:

>In a similar matchup, AlphaGo running on multiple computers won all 500 games played against other Go programs, and 77% of games played against AlphaGo running on a single computer.

But the full version of AlphaGo that runs on thousands of computers is much stronger than that, so I was mistaken.

Still, the fact that the non-distributed version is so strong even without tree search is pretty amazing. It beat all existing Go playing programs a majority of the time. And with algorithmic advances and more training it may eventually catch up to best human players.

[computerex]

I don't know about the 25% figure, but the original Alpha GO paper mentioned that their best solution is a hybrid approach between neural nets and MCTS. However, the system can beat the best Go bots out there without doing MCTS and relying solemnly on the policy/value networks, which I think is truly amazing.

[Cybiote]

According to the old paper (AlphaGo has seen significant improvement in efficiency and algorithm so this might be outdated), the distributed version with 1900 CPUs and 280 GPUs defeated the version with 48 CPU and 8 GPUs 81% of the time.

Non-distributed Alpha Go won 99% of the time versus just the value network and policy network with no rollouts. That AI was estimated as having a 2177 Elo rating, which is not very strong and much weaker than Sedol.

Even with a TPU, a human is more efficient. That neural net pair used 8 GPUs. At a generous 200 watt per GPU that's 1.6 kW, 10% of which is 160 watts. A human brain does all higher level reasoning and uses ~20 Watts. A human is not devoting 100% of its computational power on Go. It is likely just a fraction of that.

But if we look at Chess, Chess engines that run on mobile phones are possibly about or maybe slightly more efficient.

[Houshalter]

I believe I was remembering this from wikipedia:

>In a similar matchup, AlphaGo running on multiple computers won all 500 games played against other Go programs, and 77% of games played against AlphaGo running on a single computer.

But you are right, the full version running on thousands of computers is much stronger than that.

Still, the fact that the non-distributed version is so strong even without tree search is pretty amazing. With algorithmic advances and more training it may eventually catch up to best human players. It's only the first generation of deep learning based Go bots.

And I believe the policy network only takes a few milliseconds to compute a move. So even if the TPU consumes hundreds of watts at full use, it doesn't need to run at full use for long.