search

MarkZH / Genetic_Chess

An amateur attempt at breeding a chess-playing AI.
MIT License
10 stars 2 forks source link

Perform full minimax search on all depth-2 moves #107

Closed MarkZH closed 2 years ago

MarkZH commented 2 years ago

This can also be an attempt to do multi-threaded minimax with a thread pool similar to the gene pool structure for starting simultaneous games.

Might also be a good idea to create a Thread_Pool class to handle both these cases.

MarkZH commented 2 years ago

Not a Thread_Pool, but a Thread_Assembly_Line. A push_back method creates a new std::future which is put into an ordered container. This method blocks if there are no processes available to start it. The function passed to push_back() is run within a lambda that also locks and releases one of a set of mutexes to limit the number of simultaneous threads. Perhaps the mutexes can be put in a queue that blocks when something tries to grab a mutex when it is empty.

MarkZH commented 2 years ago

For that last bit, use std::counting_semaphore.

Actually, use std::counting_semaphore in the current implementation of the gene pool to see how it simplifies the simultaneous game limit.

MarkZH commented 2 years ago

Simpler still: create a structure that acts like a ticketing system. Before starting a thread, take a ticket. Have the thread release the ticket when it finishes. This will be implemented in a class that wraps a std::condition_variable for master. In c++20, that class will be deleted and replaced with a std::counting_semaphore.

MarkZH commented 2 years ago

Thread limiter implemented here: 36dd6fcbd54f0975dc8be28d7f237765a96dc46a Work ongoing in multi-threaded-minimax branch

MarkZH commented 2 years ago

Full-minimax search on depth 2 moves is very slow due to having to traverse so much more of the game tree. Reduce this to depth one, maybe with a shared Alpha/Beta values.

MarkZH commented 2 years ago

More importantly, the engine crashes while doing multi-threaded search. The location of the crash is usually random with memory locations indicating destructed data or out-of-bounds access. Investigating ...

MarkZH commented 2 years ago

All data sent to threads is now copied as of commit 2d4062b91560c9e96269aaa65b85e09f447510ee

The multi-threading was reduced to depth one moves due to slowness induced by so much searching caused by the lessened effectiveness of alpha-beta pruning due to starting with new windows with every thread. The branch is fully functional and waiting to be merged.

MarkZH commented 2 years ago

Threads can now be used in gene pools as of eda7239c6f97fcaecbfd92601254649cb8129ddc