Re: go programs

Posted by Robert J. Cordingley on Mar 14, 2016; 10:37pm
URL: http://friam.383.s1.nabble.com/go-programs-tp7587292p7587299.html

Access, for a fee, to the original Jan, 2016 Nature article on AlpahGo is at http://www.nature.com/nature/journal/v529/n7587/full/nature16961.html.  The freely available abstract says it uses deep neural networks ('value networks' and 'policy networks'), tree search and Monte Carlo algorithms. Figures and tables with more information are also freely available from http://www.nature.com/nature/journal/v529/n7587/fig_tab/nature16961_ft.html

Robert C

On 3/13/16 8:53 PM, Steve Smith wrote:

Me, I'm still stuck in the 80's... most of what I know about GO programs involves trying to solve them using cellular automata systems based on the promise of hardware implementations and other esoteric ways of doing CA computation...   Tomasso Toffolli's custom CA hardware was one promising thing that I think eventually fizzled as was our own Jim Crutchfield's analog "video feedback" CA computing concepts...   

My own favorite which I went on to do some exploratory work in was the "memoisation" work of Bill Gosper which involves generating hash tables at each scale (say 3x3, 6x6, 12x12, 24x24) cell arrays such that if "redundant" patterns occurred at any scale they could be "looked up" instead of computed.   In a 3x3 (9 cell) array, there are naturally only 512 (2^9) hash indices so the computation at that level is manageable by memoisation... while a 6x6 is 2^36 or roughly 64M entries, not quite so tractable/trivial if the distribution of possible configurations of binary CA were uniform...  which interesting GO configurations naturally are NOT.   A slight modification to this is that a binary CA is not sufficient since the states of each cell can be White/Black/Empty... so the math changes to 4^9 and 4^26,etc...

Similar attempts were made for checkers and chess which as I remember, the state space for Checkers is much larger than for Chess (surprising?) but GO... much higher (larger board!) and the depth (number of relevant moves ahead) also much higher!

I look forward to hearing what the current state of computer GO play might look like as well!

- Steve

There were stories during the expert systems episode in the 80's that some experts when debriefed in an attempt to identify their rules went on to lose faith in their own expertise and to resign from the field. Other anecdotes talked about how some experts weren't capable of expressing their expertise - such knowledge, skills & experience was referred to as 'compiled knowledge', accessible but not expressible, much like Artificial Neural Networks are. Work to address this problem has been underway since the 90's. Perhaps others here can provide an update?

Robert C

On 3/13/16 8:45 AM, Marcus Daniels wrote:

I think a deep neural network trained from self play has a subjective, and even inscrutable inner representation.  Imagine such techniques were applied to public policy decisions or medical diagnosis.   Without a linguistic component that co-evolved to describe a taken action, one could be left with robot savants that outperformed humans on crucial tasks and no one, including the robot, would have any idea why.

Sent from my iPhone

On Mar 13, 2016, at 8:01 AM, Roger Critchlow [hidden email] wrote:

I've been watching parts of the match between Lee Sedol and Alpha Go on the youtube deepmind channel.  It's quite good, they start off with a discussion of the previous game, give running commentary during the game, and audibly gasp when the progress of the game shocks them.  The post match press conferences are not to be missed, either.  It's a completely trump free zone.

But you're looking at a full day's work for each game, 6 hours and 17 minutes of video from last night's game which Lee Sedol won.  I was too tired to stay up and watch so I tuned into youtube this morning and watched the endgame.

Apparently I forwarded past the key move, #78, which a Chinese journalist, quoting a Chinese commentator, called "a God's move".  Lee Sedol replied that it was the only move he had at the time, that he had thought it would be easier to make some profit, but it was quite difficult.

So the same play is described as both creative genius and inevitable in the space of a few sentences.  Glad to know that some things will never change.

-- rec --


============================================================
FRIAM Applied Complexity Group listserv
Meets Fridays 9a-11:30 at cafe at St. John's College
to unsubscribe http://redfish.com/mailman/listinfo/friam_redfish.com

============================================================
FRIAM Applied Complexity Group listserv
Meets Fridays 9a-11:30 at cafe at St. John's College
to unsubscribe http://redfish.com/mailman/listinfo/friam_redfish.com

============================================================
FRIAM Applied Complexity Group listserv
Meets Fridays 9a-11:30 at cafe at St. John's College
to unsubscribe http://redfish.com/mailman/listinfo/friam_redfish.com

============================================================
FRIAM Applied Complexity Group listserv
Meets Fridays 9a-11:30 at cafe at St. John's College
to unsubscribe http://redfish.com/mailman/listinfo/friam_redfish.com

============================================================
FRIAM Applied Complexity Group listserv
Meets Fridays 9a-11:30 at cafe at St. John's College
to unsubscribe http://redfish.com/mailman/listinfo/friam_redfish.com