bawjensen / comp-499-independentresearch Goto Github PK

Currently the system creates many separate .csv's, and they have to be compiled by hand into an excel doc. This isn't ideal, not just for the annoyance, but excel seems to have issues with 8000 data points in a scatter plot.

What's a better approach?

Going into data-viewer folder, probably using d3.js along with a simple node server.

It's possible that some memory leaks were reintroduced with the addition of the TestSuite, and it's been a while since any sort of optimizations was performed, and it's possible that there're some large issues.

Put everything to do with actually running the net training into a separate process that is not tied into the execution of the server code, so that if one goes down the other is unaffected.

My version will merge something multiple times, for example:

row: 0 4 2 2
cmd: go right

This will collapse the 2 and 2 into a 4, and then collapse the 4 into an 8. This isn't wrong, exactly, just different.

Based on one or two parents, doesn't matter at the moment.

Currently there exists a tiny python script that runs parallel configs. This made it so that I don't need a "screen" per config. What if Python can help eliminate more issues, such as dynamically deciding how many cores to use.

Is this an actual issue or a leftover from when I was checking the before-and-after of deletion?

When going through the minimax tree, and the branch ends (that game was ended), determine whether or not it makes sense to eval the board right there, or to add some constant in to represent the fact that this board is dead?

Create some sort of set up for a queue of different runs with various configurations, that will be run one after another, possibly up to 6 at a time, if possible.

Have the directions evaluated on average score, not true minimax, as the probability may be better.

There's a problem with the case of no children of a given board. Minimax returns -inf, which breaks things.

Such as having the nets rated by their top score, allowing command line arguments to customize number of generations, nets and runs/net, etc.