rspeer / golem Goto Github PK

Golem should know when a buy possibility will end the game, and choose it if it wins or avoid it if it loses.

Looking a turn or two ahead and calculating probabilities, for the kind of endgame strategy discussed on CouncilRoom, would be fun but should come after the basics.

A failing test in golem-test.coffee shows that the bot will buy a bishop over a silver, even when its deck is already clogged full of bishops.

Somehow the model learned that Bishop*Bishop = good, and neither the actionBalance nor the weights against Bishop>3, Bishop>4, etc. are enough to overcome this.

Something probably went wrong in handling games that have already ended. Models for turns later than 13 output a constant +100 or -100.

export_decks.coffee seems to occasionally be misplacing line breaks, or perhaps writing one example in the middle of another. The result is that somewhere between 1 in 40000 and 1 in 10000 of the examples are corrupted. Is this a node.js bug, or just an unexpected effect of the way I'm doing a lot of asynchronous writes to the same file?

The workaround for now involves telling vw not to care about serious parse errors. The model picks up some meaningless features and otherwise copes.

Golem currently treats costs as integers, and doesn't take into account the fact that they may include a potion.

The buying math should be changed to keep track of a number of potions AND a number of coins.

I installed all of the dependencies (the node mongo install failed, but that seems like it's only for training anyway) and got the server running. I ran the buyborg iso client, but it didn't make any requests to the locally running server, nor did I see any part of the UI for giving recommendations.

It would be nice to have a trivial client that has a textual game state hard coded as a way to document the interface, so for example, I could hookup the cr.com game pages to make requests to the server an annotate the game with buy recommendations.

Golem plays a Hoard and has $6 to spend. It wants to buy a Gold. It doesn't recognize that it would be better to buy a Duchy and get a free Gold.

Golem plays a Goons and has $6 to spend. It wants to buy another Goons, with Goons+Copper not too far behind. If it knew it would get 1 VP more, it would certainly chose Goons+Copper.

This project is exciting to me.

I am curious about what you are trying to model from the data. Is there some code missing, or you haven't gotten there yet?

For me, there seems to be two reasonable approaches to this.

1. Try to model what a given (skilled) player will buy given the context, available money, etc. This could be fun from a social perspective, you could heavily overweight data from a particular player and then call it the rspeer-bot, or the rrenaud-bot, or the theory-bot, etc.
2. Try to estimate the probability of winning (or merely relative advantage, monotonic but possibly less meaningful than p[win|game state]), and then enumerate possible game states from a position, and use that to determine which things to buy.