I've recently been looking at Briscola, which is a traditional trick taking card game. I thought it would be relitaavely easy to produce Ai players for it, what with only ever having possible 3 moves a turn and a fixed number of turns to complete the game. I decided I would knock up a quick Monte Carlo player to see what it played like. The trouble is while I wrote it pretty quickly the AI is so slow iterating through games that I feel I must be overlooking some fundamentals of appropriate data structures to use to model the deck of cards.

Are there any good code examples of Monte Carlo based AIs for trad card games out there that I could learn from?

Thanks for your responses, would have replied sooner but had problem logging back into the site.

I had profiled my code before posting and had spotted a few obvious inefficiencies, for example due to stupidity on my part I was shuffling the deck twice when performing a play through of the Monte Carlo player and still building up logging strings despite not printing them out. Now that I have knocked out the (to me) obvious the largest amount of time by far is spent in random.shuffle(). Briscola is a trick taking game (normally 2 players) where players have a hand of three cards and each round play 1 card then draw one card blind from the common deck. So correct me if I'm wrong but each play through by the Monte Carlo player requires shuffling all the unseen cards to get a new ordering of the deck. That's unavoidable correct? I'm using Python so the shuffling algorithm is linear time and efficiently executed so I can't see much speed up there.

I think I might just have too many "speed of writing code vs speed of code execution" trade-offs in my Python programming style for this task - I should maybe do this in another language to force me to think about it differently.

PRE-EDIT: Actually, writing this post has made me think critically about the code - I may have spotted another couple of gross inefficiencies. I'll let you know how it goes.

Programming this in Python is already a problem because performance matters for Monte Carlo techniques. Do you know C++?

Programming this in Python is already a problem because performance matters for Monte Carlo techniques. Do you know C++?

I've done C and C++, although given how long since I've last used them in anger I would hesitate to say that I am proficient in either. This may be a good project to reacquaint myself with the language.

Why would you shuffle unseen cards more than once?

Finally having my first free weekend in a few months I'm almost done with my C implementation. The data structures I've come up with aren't that fancy or complicated but are pretty efficient. Once I got out of the mental dead end of the "deck of cards" actually behaving like a deck of cards with a 'top of deck' and the like I've been able to arrange the cards such that I need very few operations to save the current state and randomise the game state for the Monte Carlo playouts..

As you'd expect it's blazingly fast compared to my Python implementation, on the order of several hundred times speed up, possibly more. Once I'm done with this I'll be interested to go back to the Python code and implement the same data structures in Python - once I'm no longer allocating and freeing memory all over the place I wonder how fast I can get the Python code to run.