new challenge for Go programmers

[djhbrown]

papers.ssrn.com/sol3/papers.cfm?abstract_id=2205530

[Celebrir]

At this moment I'm quite sad I can't put as much time and effort as I would want into this

[RobertJasiek]

Even from a pure playing stronger perspective, it is not game over yet because there is no guarantee yet for always avoiding sudden entering of holes of bad play, verification by reading is missing and there is no optimisation for better score when winning the game anyway. For other AI applications, this means that there is no guarantee of unexpected bad actions, such as accidentally killing people.

As I have explained, shortly after the sacrifice squeeze in game 5 AlphaGo had a winning position. Therefore, currently one should not call the squeeze a mistake. A more cateful study of the few moves after the squeeze is necessary.

You mention several outdated principles and concepts, whose insufficiency I have explained elsewhere.

[Pippen]

I agree with Robert and I go one step forward. As long as a bot relies heavily on heuristics it's never "game over" because the very concept of heuristics includes mistakes and failures.

I'd also throw in an idea I call "escape into bigger board sizes" for I think the bigger the board the harder it gets for AI. We know AI got to dan level on 9x9 first when it was still kyu level on 19x19. If this logic is linear it'd mean that the best humans would still clearly beat any AI on 21x21 and that even decent amateurs could beat machines on 25x25. I think one could escape up to 99x99 boards - by agreeing to play only certain segments at once and to play the complete game within days. Thereby we might also be able to witness the very first Go game that produces some kind of meaningful picure as a collateral of playing.

[jeromie]

Your ideas are intriguing, but I'm not sure how well it would work. It seems that AlphaGo is strong precisely because it does not set intermediate goals for itself. This is what allows the "creative" moves and surprising tenukis. The fovea may allow for optimal local play, but what if the optimal play is not local?

[djhbrown]

When we move beyond the limited world of a Go board to ask an AI to do intelligent things in the real world, i personally don't think CNNs with or without Monte-Carlo simulation are going to be enough and that some kind of hierarchical reasoning system is necessary.

Two lines of research that i find promising are Winston's work on story understanding and Wolfram's work on problem solving, both of which use hierarchical ontologies.

btw Wolfram Alpha has nothing to do with Alphago.

http://cbmm.mit.edu/video/patrick-winst ... ding-story
https://www.wolframalpha.com/

[topazg]

tbh, I disagree that there's little value in simply improving the playing strength of the engine, and I would be surprised if AI Go programmers feel that way either. Chess had players finally beat arguably the world's strongest player in 1997, and have continually improved heuristics, ideas, concepts, and pure calculating power of engines ever since. Back then Deep Blue was around the strong 2800 mark and non-custom-hardware engines probably around 2700, now my laptop can download open source software running over 3300.

They still aren't good at saying "hey, you shouldn't go here yet.. I mean, it's a good move but you need to prepare it with this and that so you're threatening this and that" ... they're just stronger, and that seems like a valid end all by itself. Top grandmasters now routinely rely on them to test out their home prepared opening lines for traps and unforeseen complications.

[jeromie]

I think there is value in making the strongest engine possible. There is also value in making an AI that can communicate its decisions in terms a human can understand. The question is whether the latter can be made strong enough for its decisions to be interesting. I hope someone takes up the challenge of implementing such an engine!

I think both approaches will have utility beyond go, too. There are times where we just need the best possible decision to be made. There are other times where we need a system that can train people or augment human reasoning. The latter approach probably engenders more confidence in its decisions, too. It will certainly be interesting to see how various forms of AI are incorporated in our daily lives.