Lines of Action - Using TD+NN for a deterministic perfect information board game

Alistair_Hutton · 2007-02-23T14:24:56

I've recentely become interesetd in the board game Lines of Action and have been poking away at creating some evaluation functions for the game. I've simultaneously become interested in reinforcement learning and thought that this game woukld be a good test bed for practising. One thing I've seen mentioned in the literature is the introduction of random moves to prevent the NN getting stuck in a loop. I'm wondering if there are any good guideline for introducing the random moves. Should it be a fixed percentage of moves or should the randomossity ramp up if the game is going on to long Two totally random players (not even checking to see if the next moves leads to a winning/losing state) facing off against each other take about 120 moves +- 30 all together to play a game to completion. A very basic hand crafted evaluation function with a half-ply search wins against the random player in about 23 moves.

Artificial Intelligence Programming

Started by Alistair_Hutton February 21, 2007 10:28 AM

9 comments, last by xEricx 17 years, 11 months ago

xEricx

573

February 23, 2007 02:24 PM

Quote:
Original post by Timkin
Quote:
Original post by InnocuousFox
Related question: Has anyone ever put a bunch of NNs through a generational process like GAs? Talk about over the top! :-D

Optimisation of parameters in ANNs using GAs has certainly been done to death. Structural adaptation is less prevalent in the literature.

Isn't NEAT exactly that?

[size="1"]boost-loki-game ai by example-ai wisdom 3

Lines of Action - Using TD+NN for a deterministic perfect information board game

This topic is closed to new replies.

Popular Topics

Recommended Tutorials

Lines of Action - Using TD+NN for a deterministic perfect information board game

This topic is closed to new replies.

Popular Topics

Recommended Tutorials

Reticulating splines