A bot to play solitaire in messenger.
Compiling: make
Running: make run
There are a few components to it:
- JNI interface with the
Robotclass to capture screenshots, control the mouse and keyboards. - Using template matching in OpenCV, new cards are identified. That is converted into a game state.
- The bot "cheats" by going through all the cards in the draw pile in the beginning.
- The game state is purely markovian, hence a function
move_t solve(state_t)can decide the best move. - As soon as the solver encounters a fixed point (that is
execute(state, solve(state)) == state), the program terminates.
I had some problems getting everything running with a headless JVM (i believe it is some problems with X server or something along that line), hence, I ran everything in a java proces, calling into JNI method on every frame.
- How easy is it to get it running on my own machine?
I am glad you are interested in this project! You will need to tune some
vision parameters in test/vision.hpp depending on your machine, and where
you decide to position your window. I wish to automate this process, at
some point.
1. Why don't you machine learning this?
Much wow.
2. Why don't you treat this as a markovian decision problem and run some TD-learning algorithms on this?
I didn't really understand reinforcement learning (I still don't), but from what I can gather, this is a non-deterministic purely markovian decision problem with perfect information. This gives it really nice properties:
- randomness of the game prevents it from stucking in local optima
- it has a finite discrete action space
7 + ((7 + 1) * 4) + 1 + 4
so, I'd expect even the simplest greedy TD-learning algorithm like those in TD-backgammon work.
there has been quite a bit of previous work in this -- a simple search in google scholars yield some interesting results. A fun experiment to try "learning" to see if a policy to "remember" what it has seen in the draw pile without explicitly adding it in the state.
3. Why did you have to partially implement this in C++, rather than just use Java on the whole stack?
I don't want to java.
4. The code looks like a mess. Wh....
I agree.
5. Why number_t rather than rank_t?
I am not a native english speaker. I didn't know that it is called rank until I shared the project with my friends.
MIT