The Real Test for Komodo 4

[mcostalba]

However, there is a significant problem. I changed my idea seriously. I wanted to make it faster and stronger. But not so that Sting could be stronger than Stockfish.

Please Marek don't be shy, feel free to make it stronger than Stockfish


P.S: Not to discourage you, I like your work, but I think this is more easy to say than to do.

[Uri Blass]

You do not need to suddenly switch away from random moves to get senior master level if you use probabilities.

probability of 50% for random moves is clearly beginner level but
I believe that probability of 1% for random moves is enough to score more than 50% against master level and the main disadvantage is that it is going to be unstable and people are going to complain that in one game it plays like a super GM because it is lucky to play no random moves and play like the real komodo when in another game it simply miss mate in 1.

[fern]

Ok Ok Don, you twisted my arm.
I will rey games from 0 level up to...whar? we will seee....

Fern

[Don]

You do not need to suddenly switch away from random moves to get senior master level if you use probabilities.

probability of 50% for random moves is clearly beginner level but
I believe that probability of 1% for random moves is enough to score more than 50% against master level and the main disadvantage is that it is going to be unstable and people are going to complain that in one game it plays like a super GM because it is lucky to play no random moves and play like the real komodo when in another game it simply miss mate in 1.

So you are suggesting that the program plays at 3000 level but once in a while throws in a random move, but that is not a practical solution. I think even a senior master would find it annoying that the program plays like a 3,000 and once per game plays Kf1 at a completely inappropriate time. Even if the move looks like a blunder it should at least be a "plausible" blunder. Even a weak player would not likely move his queen to a square where 5 different pieces could capture it for example. The queen might get attacked by one piece and not noticed, but nothing more.

It has to involve the depth of the search because that automatically makes weak moves plausible, even if weak. There are a number of solutions but they have to be carefully considered, such as having a very short quies search, randomizing the evaluation function, etc. I don't like random moves because I believe even the weakest of players do not play randomly, even if it looks like that to strong players.

Probably the weakest level that is reasonable is a 1 ply search with highly randomized evaluation and a crippled quies search that does not consider checks and randomly prunes even non-losing captures or something like that. The randomness for pruning captures and evaluation could be smoothly phased out.

I basically want to create a "plausible" weak player. I know I won't be able to this perfect.

[Don]

Ok Ok Don, you twisted my arm.
I will rey games from 0 level up to...whar? we will seee....

Fern

Good! Let me know how it works for you. The idea is to decide in advance what time control to play and stick with that, adjusting the level upwards until you have an even match. Report back to me (it can be private) on your experiences, either good or bad and advise me.

[fern]

A weak player is basically one that is not good at calculating moves, so he fall in traps, do not see threaths, calculate wrongly or at least calculates OK BUT evaluate badly the ouput.
I know that very well because I am one of those players.
I suppose the nearest to all that, as we talked long ago, is deepness of calculation. Dedicated comps were winable because they calculated no more than 6-7 plys in middle game, which leaves them more or less inside the territory of mediocre players.

Fern