So its been quite difficult for me, John, or Carlos to get Stella AI running on dice, or any of our laptops. A big part of this is a number of dependencies on external libraries, and silly differences in naming of libraries etc etc etc…
Since Michael was interested in using a more honest (thats my word, not his) image segmenter that just cuts out bounding boxes around colors, I decided to change course a bit. This approach will probably give us poorer features, but if we can get learning working with it, it will be more impressive and general for other games later.
Stella AI is written in C++ (from what I understand), a language I won’t use. Another issue that seems to prevent Stella AI from playing nicely seems to be the use of Lua, which I don’t know anything about (apparently, there are scripting languages that aren’t named Python). At any rate, I found a version of Stella implemented in Java (JStella).
Just today I downloaded the JStella source, added a few lines to the source, and am dumping out video memory information. The picture above is super crude because video memory just has values of indices to the palette, and I just dumped those numbers out and drew a chart in Excel. Regardless, you can see whats going on. Yes I know it seems oddly tall, not sure why the buffer is like that, but its there; if i was going out of the buffer id get an out of bounds exception, of course.
So, I’m going to go ahead with this. JStella is very nice because it needs no libraries aside from the stanard Java libraries, which should keep headaches to a minimum, which is pretty important in a project that has so many different pieces to keep track of.