Question

[andyjr5566]

How long does it take for the bot to be will trained?

[grschafer]

Here are videos of its performance:

• completely untrained: https://youtu.be/ew_O0HdWd5Q
• after 3 hours of training on a mediocre consumer desktop: https://youtu.be/7wey-S6H38E

From my original slides for this project, here's a chart of its ability to make matches as training progresses:

[andyjr5566]

Thanks,
Another Q, what environment features did you feed as input to the network?

[grschafer]

My ML knowledge is fuzzy these days, so apologies if I use any terms incorrectly. This implementation uses Q-learning, so it's backed by a giant table (rather than a network), and the table is indexed by state-action pairs. So, the current state of the board is the only input. I simplified the state space by only considering a 4x4 board of a single color, as shown below:

Then, I looped over each color and looped over each arrangement of a 4x4 window on the normal 8x8 bejeweled board to find the highest-reward action available.

[daedus]

can i train it and then use it on other games ?