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?

"Iterating through games"? Didn't you profile the program and/or try different problem sizes to measure which parts of your training are slow? For example, variant games with 8 AI players and 5 turns should be faster than games with 2 AI players and 20 turns if you have unneeded overhead per turn, slower if you have unneeded overhead per player, and equally expensive if the cost of each of 40 AI player moves is the same in both cases.
 
Can you explain what you are doing in more detail? What state and inputs do your AI players have?

Use a profiler to figure out what parts of your code are slow, then you can try different implementations. If you can't figure out why some part is slow, perhaps you can describe to us how you implemented it, in case there is something obviously wrong.

One general rule is to avoid dynamic memory allocation. Depending on the language you are using, this might be tricky to do. But with a bit of discipline, you'll realize that you don't really need it.

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++?

You need a deck shuffle at most once per game, and it cannot plausibly cost more than the 1 or more random numbers per turn a "Monte Carlo" player presumably needs to choose a move (40 cards, 40 turns). While playing, the order of the cards remain fixed.

You can even recycle the same shuffling of the deck many times with different random moves in each playthrough; depending on how do you evolve the AI players there should be effective ways to avoid overfitting.

What representation of game state, of available moves and of their own behaviour do your AI players employ? Reading between the lines I suspect horribly simplistic and inefficient methods, but with your sketchy descriptions I cannot be sure.

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.

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.

Why would you shuffle unseen cards more than once? They are already shuffled, so that the next card is unknown to the player. I see nothing more to be gained by shuffling it again.

Why would you shuffle unseen cards more than once?

You need to shuffle them once per game. If you are using Monte Carlo methods, you need to simulate many random games to accumulate statistics and you make decisions from the statistics.

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.