Advertisement
Guest User

Untitled

a guest
Dec 14th, 2019
139
0
Never
Not a member of Pastebin yet? Sign Up, it unlocks many cool features!
text 3.02 KB | None | 0 0
  1. (17:52:07) 351062: Holly shit RoboStac, C++ translation alone ?
  2. (17:52:13) RoboStac: yep
  3. (17:52:18) Neumann: w t f
  4. (17:52:18) RoboStac: it's the same nn weights
  5. (17:52:34) Neumann: There it is, another multi where NN is the only way to win :D
  6. (17:52:35) Neumann: gz
  7. (17:53:16) RoboStac: turns out nn's are very easy to optimise with avx
  8. (17:53:33) RoboStac: getting about 7x as many predictions per turn
  9. (17:54:32) Neumann: So you made AlphaGo and run it on a AVX-MCTS ?
  10. (17:54:35) Neumann: geez
  11. (18:03:07) cdi-18290-Karim-Khenchoucha est désormais connu sous le nom de falconika
  12. (18:19:47) DJIronScone_627a est désormais connu sous le nom de ME262
  13. (18:53:33) 351062: Could you describe what you coded to achieve that ? A local arena ? The soft to train the NN (or did you use python libraries for that ?)
  14. (18:53:37) Vous êtes désormais connu sous le nom de Neumann
  15. (19:00:20) RoboStac: the data is generated by playing lots of games with my localarena and saving (state,policy,value) for every turn (state is what goes into the nn, policy is normalised visit count per move and value is set to 1 for win, -1 for loss, 0 for draw after the game ends)
  16. (19:00:50) RoboStac: I then load that data into python and train it with tensorflow because I'm not crazy enough to try and write my own nn training
  17. (19:01:22) RoboStac: the python dumps out the weights into the local arena and runs another batch
  18. (19:01:37) RoboStac: it retrains every 500 games
  19. (19:02:58) eulerscheZahl: why can't you just be a normal guy, giving us the illusion of having a chance against you?
  20. (19:03:46) Neumann: "the python dumps out the weights into the local arena" so your local AI is able to read weights in a another file ? Or a database ? Or you just dump the weights, sed in your AI src and recompile ?
  21. (19:04:21) Natendrtfm: does Javascript a good languages for codingames or not ?
  22. (19:04:22) RoboStac: yeah, the arena reads the weights from a file
  23. (19:04:48) Neumann: You mean your AI your arena is just the referee right ?
  24. (19:06:38) RoboStac: it's referee + ai code that plays games locally
  25. (19:07:54) Neumann: All in the same code ?
  26. (19:08:23) RoboStac: yeah, I didn't want to mess around launching lots of subprocesses
  27. (19:08:58) Neumann: Is that what you would do in a "regular" contest for local benchmarks ?
  28. (19:09:25) RoboStac: nah, I'd normally use brutal tester
  29. (19:09:45) Neumann: Isn't that too slow ?
  30. (19:10:06) Neumann: reCurse, Agade, etc seem to all recode a C++ local arena
  31. (19:10:37) RoboStac: it's slower, but during a contest I'm not sure it'd make that much diffference
  32. (19:10:44) Neumann: Ok
  33. (19:10:56) RoboStac: it does here as I'm trying to play a lot more games
  34. (19:11:16) eulerscheZahl: doesn't that also depend on the game? for some like Oware you can almost certainly say that playing stronger vs yourself means playing stronger in the arena toofor other games there are too many different viable strategies
  35. (19:11:21) RoboStac: I think the current oware submit was after about 500k games
Advertisement
Add Comment
Please, Sign In to add comment
Advertisement