« first day (1516 days earlier)   

vzn
2:45 AM
in theory salon, 2 mins ago, by vzn
hi all thx for all the great chats. great to see this & some other SE chat rooms livening up recently. inspired to come up with some ideas from 3yr experience/ participation in misc SE chat rooms. suggestions/ feedback welcomecool chatting session tips / stackexchange profile tips
 
 
6 hours later…
8:38 AM
@Gilles do you know offhand if gcf is solvable in polynomial time?
came across your answer for cs.stackexchange.com/questions/1447/…
trying to find answer to that...
 
8:59 AM
nvm, found it :)
 
 
6 hours later…
2:54 PM
@vzn very long post...
 
vzn
@EvilJS it was built up incrementally over ~3yr and the two sections cited are new yesterday.
 
@vzn ok, I have seen it earlier, especially part where you use small letters... It is just long-ish as one piece.
@vzn anyway, I was poking (proper word?) to ask you about ideas - there are a lot of games, chess, go, good engines etc.
But how can we measure current state? I mean heuristics, tables, calc forward moves - but even with heuristics we use calculated earlier partial results, so it begs for horizon effect.
So... do you know any games where the order does not matter? Like hex, but easier, or games where current state is very good description of who wins?
 
vzn
3:19 PM
@EvilJS current state of what? game engines/ intelligence? that just happened in big way last mo with alphago competition etc
@EvilJS youre interested in the AI of game scoring?
 
Current state as in board - you move C, I do F, you do H and this is state.
 
vzn
deep NNs can now "score" games better than humans. there is some application of them to diverse games eg chess. in fact (well concealed trivia here) it appears a chess reinforcement learning program may have been the precursor to alphago.
 
Deep NN, isn't it just NN with brand new name?
 
vzn
so basically deep NNs are state of the art, and, as far as anyone knows, the answer to everything.
@EvilJS lol, re gartner curve, there is some major hype, and some major nonhype.
also nearly all games less complex than go have human-beating algorithms, some long ago. etc
 
Ok, so it is one way to go, I was rather looking for some game where calculating who is winning is doable from the current board state, so program wouldn't be too involving and people playing would show what they are doing
 
vzn
3:24 PM
why not chess? it fits that criteria. what do you mean "people playing would show what they are doing"?
are you talking about a classroom exercise?
 
@vzn I meant to know who is winning, but we do not have perfect strategy for chess yet, so counting pieces and giving weights is far too bad. No aftermatch for chess game - too long experiment, I am looking for easier game.
 
vzn
what do you mean "aftermatch"?
"easier game"— checkers?
 
follow up experiment, the next in the series.
 
vzn
sorry not following
 
Yes, I was thinging exactly about checkers, but did my time, so... At least people I checked were not using any strategy.
My bad, I tried playing chess game experiment - now I am after other game, this is follow up, ok?
I haven't mentioned, but checkers - I haven't found people who played checkers, it was almost random
 
vzn
3:31 PM
what went wrong with the chess experiment?
 
It was taking too long to present possible moves and too long to generate them
It was exhausting for player - time consuming, so it was not good to perform many games. Also too hard to definitively tell strategy
I did Chess, checkers, will try Hex, will not try go.
Checkers - draughts, different rules, but the only pattern is people reduced the space, with rule that you have to beat, so... initial phase is like get rid of 3/4 ply and then start playing...
Optimum is more than 25 games played per person, I could not do more than 5 chess games in one day - this is the problem. Very good idea is "hey that is my fish", but too heavy for AI
So I am looking for game, which is easier for computer, but involving for human.
And there is the thing, like with sokonan games, it is easy to order games by number of moves, but it is not correlated with difficulty.
I hope you know what I mean?
Oh, putting it another way - number of lookahead moves in chess is not enough to control difficulty.
 
4:18 PM
Hi @sandy
Changing the subject, there is nice question which is really helpful, at least outcome would be awesome. Are there any insights?
 
4:38 PM
@vzn I do not do in classroom anything connected with private research, nothing even slightly helpful for me, they have to learn certain topics, I do different things outside, those things do not collide.
 
vzn
5:04 PM
@EvilJS huh? just go back and read your own words sometimes & see if they all make sense...?
hey so how big is your class anyway? why did you say the teacher was unavailable, what happened there?
@EvilJS are you saying chess (even without software involved) requires too much mental exertion by subjects for your experiments? anyway even halfway decent chess code can generate moves quickly, so it would only be somewhat amateur or poorly written code that takes too long to evaluate positions...
 
@vzn tried that, but got lost ;)
@vzn about 30 students, he had car accident, alive, recovering, will be fine.
@vzn I mean that showing stimuli is involving, time consuming. No, poorly written ai player is fast and works good enough. Better player is not that much slower, the main problem is showing results and that tweaking difficulty level is hard - increase in lookahead, some constants, preference of one ply (like knight) over another (like rook) is not easy to manipulate difficulty - for human to give him involving enough game, but not that hard enough, so he could win.
If you change 12 levels depth to 13 levels, this is not exactly linear increase in difficulty, also some moves may be expected by human player for different reasons than his lookahead. So I cannot even guarantee that separate games for the same player are at the same level, so the results are tainted.
 

« first day (1516 days earlier)