3

I'm trying to get a gauge on just how big the programs and databases are these automata. I understand that this is a changing number, particularly in regard to Machine Learning.

Q: How large was Deep Blue when it beat Gary Kasparov?

Q: How big was AlphaGo when it beat Lee Sedol?

DukeZhou
  • 6,227
  • 5
  • 25
  • 53
  • 1
    Let's get the concepts correct. AlphaGo uses Monte-Carlo Tree Search, it doesn't prune. Also, we don't prune a database, we search from it. We prune a search tree, like the alpha-beta algorithm. – SmallChess Mar 31 '17 at 00:13
  • @StudentT thanks for the correction. (I've amended the question to remove the problematic language.) – DukeZhou Apr 01 '17 at 19:48

1 Answers1

2

AlphaGo used data from the KGS Go Server, which had 160,000 games and 29 million board/next-move pairs. But crucially, after it was trained on the dataset, AlphaGo was trained through self-play, so its competence shouldn't be measured strictly in terms of its database.

I'm not 100% sure how Deep Blue worked, but I think it was a mix of 1. a "book" of opening theory 2. explicitly coded board evaluation functions 3. a "book" of endgames. So there isn't a "database" in your traditional "ML by big data" sense. But in any case, I would assume the bulk of the work is done by the evaluation function, so again its strength cannot be measured in terms of if its database.

k.c. sayz 'k.c sayz'
  • 2,091
  • 10
  • 26
  • Thanks for answering! I should clarify that my project involves non-networked automata, restricted by the resources of a single device, and the allocated volume of the gamespace (the app) which ideally is trivial. (The goal is semi-strong AI that for a metagame with complexity < Chess to > Go that runs on a mobile phone. Metagame here is defined by a set of core mechanics that can be extended in numerous ways, including board configuration and value adjustments which do not require additional mechanics, but alter equilibria.) Thus I'm interested in raw bytes. – DukeZhou Apr 22 '17 at 04:01
  • Great point about Deep Blue--I now want to ask a question about the size of the programs, which would include the evaluation functions, which I assume are fairly weighty in the IBM case, from what I know about that program and it's development. Game books I definitely consider database content--it's still data stored in memory for access. – DukeZhou Apr 22 '17 at 04:02
  • 1
    That Im not sure about. I recall IBM was very secretive about this (going as far to disassembling the machine not long after the match), and a quick search shows nothing regarding the specifics of the algorithm asides from the very high level concept. Regarding AlphaGo, though Google has been quite open with the research itself, I don't think they have released any code (at least not the trained version). – k.c. sayz 'k.c sayz' Apr 22 '17 at 04:08
  • 1
    Off handed note. I don't recall from where I've read this, but the fact that chess programs are better than GMs now has more to do with better computing power than better algorithms. – k.c. sayz 'k.c sayz' Apr 22 '17 at 04:13