Polyglot Opening Book of Random Moves?

[Steve Maughan]

Has anyone created a Polyglot opening book of the first "x" random moves? It's relatively simple to do — just do a perft for "x" depth and add each position to the opening book. Just wondered if it's already been done.

Steve

[MOBMAT]

That would created a very large book with very little depth. After the first 3 full moves there are almost 120,000,000 positions. doesn't seem it would gain you very much.

If you are developing your own engine, you could add a command to play random moves for the next X numbers of plies.

[Steve Maughan]

The use-case is to create a large number of games with a huge variety of positions that will be used to train a neural network.

Steve

[hgm]

Why do this through a book? Just let the engine pick the first N moves randomly. In Fairy-Max I add a random value to the score in the root for the first N moves to create diversity. But there I limited it to 50cP, so that it would not blunder away material because of it. But if you would make it asymptotically large, you would get pure random play.

[chrisw]

The use-case is to create a large number of games with a huge variety of positions that will be used to train a neural network.

Steve

If all you want are labelled positions you could write a self-play with random book for any UCI engine in Python in about 20 lines of code.

[Steve Maughan]

Hi Chris — that looks like the best option. Python Chess looks excellent.

— Steve

[Steve Maughan]

Why do this through a book? Just let the engine pick the first N moves randomly. In Fairy-Max I add a random value to the score in the root for the first N moves to create diversity. But there I limited it to 50cP, so that it would not blunder away material because of it. But if you would make it asymptotically large, you would get pure random play.

Hi H.G. — that'll work for Maverick but what about the other side?

— Steve

[JoAnnP38]

Why do this through a book? Just let the engine pick the first N moves randomly. In Fairy-Max I add a random value to the score in the root for the first N moves to create diversity. But there I limited it to 50cP, so that it would not blunder away material because of it. But if you would make it asymptotically large, you would get pure random play.

Hi H.G. — that'll work for Maverick but what about the other side?

— Steve

So, you are not doing self-play to generate your positions? If so, then random move scores in the engine seems like the way to go. If you are wanting to use another engine that you don't control the source to then I better understand your dilemma. However, I don't know enough to say whether that's necessary or not.

[hgm]

Hi Chris — that looks like the best option. Python Chess looks excellent.

— Steve

Well, there isn't any need to program anything yourself for this. You can just use an engine like Brutus Random to play a zillion games, and convert the resulting PGN to a pook using Polyglot.

[chrisw]

Hi Chris — that looks like the best option. Python Chess looks excellent.

— Steve

Well, there isn't any need to program anything yourself for this. You can just use an engine like Brutus Random to play a zillion games, and convert the resulting PGN to a pook using Polyglot.

Well, there isn’t any need to go downloading external stuff, working out how to get it to work, clogging up your HD with humungous PGN data and humungous book binaries for this. Just use Python, keep the whole thing under your own control (like add filters in the future if desired), generate depth programmable random moves on the fly and dump the resulting data as PGN or EPD.