Hacker News Reader

Comment by mcbuilder | Hacker News Reader

Would not Alpha Go Zero and related algorithms give us a good insight into the question about what openings alien's might play? Isn't it really tackling the game from first (or zeroth if you will) principles, just simulated games without the historical baggage coming from human study. It has also considerably changed our own knowledge of the opening.

View on HN

I don’t think so. The go that AlphaGo plays is one very specific type of ”alien go” that exploits the strengths of the computing substrate it thinks on. Alien entities that are still evolved via natural selection would probably play go differently from humans based on their own biological and cultural idiosyncracies, but I doubt it would on average be closer to how a Monte Carlo tree search running on transistors plays rather than in another direction entirely.

This is a good point. While Alpha Zero etc. are much stronger than humans, they may have weak areas. Actually we know one: ladders.

Ladder is a fundamental technique that is easy for humans but hard for both monte carlo and neural network static analysis. It is plausible that an equally strong but more ladder-aware player would have very different tactics from the current best computer players.

The original AlphaGo actually had a hand-crafted ladder feature input layer. It was removed in Alpha Zero. We know from Leela Zero and ELF Go game records that they don’t understand ladders very well, so likely Alpha Zero didn’t understand them either.

Actually if anyone here follows current go bot development, I would like to know if they currently have some specific ladder features to address this weakness?

Are you implying alien societies wouldn’t (don’t?!) have their own historical baggage?

No, but I didn't make it clear in my question that I was assuming that the aliens would devise a similar Alpha Zero algorithm to that we have. The concepts of self play, reinforcement learning, and MCTS are obviously more universal than the exact year komi was introduced. I was more asking if using an "optimal" algorithm would reduce historical baggage in both aliens and humans, and lead to a similar convergence in terms of good go openings. However, even our choice of algorithms may be based result of historical events, and can we really conclude we've ever really found the optimum.

Reply on news.ycombinator.com