Can the bots help analyse a game? Einstein forum

19 replies. Last post: 2020-04-18

Reply to this topic Return to forum

Can the bots help analyse a game?
  • Carroll ★ at 2013-03-26

    We had a long discussion Marius and I about move 258, and the fact to take one's own pawns or not. It has been long but we could not agree on a definitive answer as we lack tools to get real answers.
    [game;id:1509069;move:258;title:best move if blue get a 3?],

    Marius agrees that I ask here.I argued that Blue should take its 4 with a 3, not my 3. Marius said that :
    “I’d have taken it with a 3, yes. And with a 1, obviously. It would probably allow your 1 to move forward, but also make it vulnerable. With a 4 or 5, I’d take your 1 with my 4. With a 6, I’d take my 4. All decent options (with the 3 the least good of them). But with a 2, there is very little interesting for me to do, and your probability of winning increases dramatically…”
    Can a bot or a good player give his opinion about the value of these moves and eventually the decomposition on future options as well as the uncertainty of the evaluation?

  • YHW at 2013-03-27

    Here are some analyzes of naive_child_c:

    - 1: The program sees blue slightly favored by capturing the red 3 with an estimated chance of 52%. (42% for capturing the own 3).
    - 2: the Monte Carlo algorithm prefers the upwards move of/for the blue 2 with an estimated chance of 32%. The chance for the diagonal move is slightly lower.
    - 3: capturing the own 4: 42%, diagonal approx. 41%, capturing the red 3: 36%
    - 4: diagonal - capturing the red 1: 67-68%; c3-c2: 35%
    - 6: capturing the own 4 or 3: 43% both

    I repeated the tests with different search depths (the MC part started first after the 3rd, 4th or 5th move) 2 different heuristics as well as different search depths for the heuristics. The preferred moves and the results were always equal (+- 1-2%) probably due to the internal 5-stone endgame database.

  • Carroll ★ at 2013-03-28

    Thanks a lot YHW for this analysis!

    The moves after getting a 3 were quite close, no wonder we did mot manage to agree! I was right to be confident moving my 3 here as it feared only the 1.

    May I ask you to do the analysis after the three moves for Red (42%,41%,36%) to see how these odds break down?

  • Carroll ★ at 2013-03-28

    A second thought about taking ones own pawns, is it included in NaiveChild's evaluation function that to have less pawns or faster ones is better?
    Do you consider it could skew up the results or is the depth you used big enough to cancel this out?

  • YHW at 2013-04-02

    Sorry, but I would have to rotate the board and enter each following combination and roll by hand. That's currently too much work, maybe later. ;)

    The evaluation function of the program is consisting of 2 terms, the mobility of a stone as well as a second value for the position of stone.
    If the value for the position of a stone would increase linearly in direction to the goal, maybe 1,2,3,4, it would not be a difference if a stone is far away from the goal or closer, the difference between current and following position would be equal. If the gradient for the position value in direction to the goal increases with a higher potency than 1, the stone closer to the goal is preferred. In this case the program is creating own attacking stones as well as defending stones, it starts rushing, that's much better!
    The values and the relation between mobility and position value are optimized in billion of games as well as the gradient in direction to the goal for the position.

    In a second evaluation function the value for the position gets 'attack-' or 'defense-benefits' (also optimezed in billion of games) depending on the stone positions of the opponent based on my own experience as a player. This heuristics plays 54% (lowest calculation depth) against that version without my own rules. Unfortunately is the calculation time three times longer, that's why the program uses that version not anymore on LG, but the other three parameters (mobility, position value, gradient) are still adjustable and changes the preferred move sometimes significantly - but not at the postion above. I tested 2 different settings, there is no big difference.

  • Carroll ★ at 2013-04-02

    Thx YHW, don't take more trouble to do that, it is not worth it.

    I'm amazed to hear that with good heuristics and plenty of power one could have a 4% edge against your bot, I thought it played near perfection!

    Could you tell the values it converged on for the gradient or is it a secret? Also I don't understand why a linear gradient makes the positions of pawns indifferent to their advance, could you elaborate? (If a pawn to a distance 1 gets a better evaluation than keeping it at distance 2, what will counter-balance this?).

  • YHW at 2013-04-03

    Haha, it would be great, but the bot is far away from playing perfectly…

    The heuristic subtracts the values of the evaluation function of opponent's stones from that one of its own stone-positions. If 2 players would have any similar stone combination on board as well as exactly the same distances to the goal (calculated by the sum of all six stones) the difference between both positions would be always only 1 (without capturing) in the following move at a gradient of 1 - independent of the stone you're choosing. If the gradient is higher than 1, stones in front are preferred because the difference gets higher in comparison to moving a stone from a starting position.
    I'm not sure at the moment but I think the heuristic with the currently used gradient wins 64% of all games at lowest thinking depth against that version with a gradient of 1. That odd is increasing significantly for higher calculation depths.
    The position values for a single stone are increasing with a potency of 2.5 - 2.6 multiplied with a basic value, means: (basic value x (5 - distance to the goal))^2.5. The total function is more complex, consisting of additional benefits depending of opponent's stone positions as well as a term considering the mobility of a stone.

  • Carroll ★ at 2013-04-17

    A strange move from naive_child_c:

    [game;id:1508994;move:187;title:Would you play a childish move here?]

    The child played 6.d1-c1, how can this be better than c1-c2?

  • Carroll ★ at 2013-04-17

    typo he played c1-d1…

  • Carroll ★ at 2013-04-17

    Oh yes it is a game in progress, but I don't think Ray or Naive_child object, as I don't think there is too much hidden strategy in EWN, or is there?

  • Marius Halsor at 2013-04-17

    The game will probably end soon anyway, and I still feel more comfortable waiting with a comment until it's over - as a matter of principle :-)

  • naive_child_c at 2013-04-17

    Do you mean 188.6/c1-b1 ? This move is not that strange as it seems. What I can say is, the program sees the chance to win the game for this move 10% higher in comparison to both others. I would take the position if I would be asked ;)

  • YHW at 2013-04-17

    Wrong login….

  • YHW at 2013-04-17

    By the way, if someone asks me whether I use the program to analyze my own games during a match I can say definitely no ! On the other hand side, it happens sometimes that I'm doing my moves and start wondering after the 60th click why I've got so many open games without realizing that I'm playing for the program….

  • Force majeure at 2020-04-15

    According to the bot that I've just created (that in my opinion is pretty strong, it shall be on LG site in the near future):1: capturing the red 3 with an estimated chance of 53%. (47% for capturing the own 3).3: capturing the own 4: 54%, diagonal approx. 56%, capturing the red 3: 35%4: diagonal – capturing the red 1: 68%; c3-c2: 36%6: capturing the own 3: 42%, own 4: 44%

  • Force majeure at 2020-04-15

    I can't figure out edditing on this forum:

    • 1: capturing the red 3 with an estimated chance of 53%. (47% for capturing the own 3).
    • 3: capturing the own 4: 54%, diagonal approx. 56%, capturing the red 3: 35%
    • 4: diagonal – capturing the red 1: 68%; c3-c2: 36%
    • 6: capturing the own 3: 42%, own 4: 44%
  • struggler at 2020-04-15

    Nice results, Force, interestingly they are quite the same as the ones posted by YHW (or computed by his bot). Waiting for your bot!

    BTW congrats for reviving a 7-year-old topic :D

  • Force majeure at 2020-04-18

    Thanks. His name is DiceRoller, and he is willing to challenge anyone :)

  • Force majeure at 2020-04-18

    Thanks. His name is DiceRoller, and he is willing to challenge anyone :)

Return to forum

Reply to this topic