Interactive Lode Runner Evolution

[schrum2]

The code for interactively evolving Lode Runner levels has some odd behavior since it was primarily developed to evolve Mario levels. Specifically, there seems to be a common situation where the width of the level is mutated to be at or around 16, the width of Mario levels. Although Lode Runner should allow for mutation of the level size, it should not be weirdly biased in this way.

I'm pretty sure that for Mario, I set things up so that there were bounds on how big/small the level could get, but I think I also restricted the precise width dimensions that were possible, since certain irregular widths actually cause the unconditional diffusion model to crash. Keep this in mind if you notice any weirdness in the code. Let me know what you see, and we'll talk about how to fix it.

[williamsr03]

Lode Runner levels now seem to always be more square with the smallest width I got after a couple generations was 16 (only 1 or 2 instead of all of them) and the largest being 58

[williamsr03]

Running the following command line produces the following picture results for Lode Runner using the evolve_interactive_conditional_diffusion. Th error I was getting about the channels is no longer there since the channels are now 8 instead of 13 (as they should be). I believe that all of the evolve_interactives can now be used with Lode Runner, but I'll quickly double check them to be sure before closing this issue. (Pictures look bad since it is checkpoint 0, once the diffusion model is finished running, I will try again using the whole model to compare)

python evolve_interactive_conditional_diffusion.py --model_path LR-conditional-regular2\checkpoint-0 --game LR

[williamsr03]

The evolve_interactive_unconditional seems to work as expected with the Lode Runner data.

Howeveer, with the evolve_interactive_wgan, it seems to create the first batch of levels fine but after evolving it seems to always create levels that are much longer than 32 tiles (my guess is 128)

[williamsr03]

After checking with a Mario wgan dataset, the same thing happens, the levels width say that they are only 16-30 but visually they seem to be at least 64, if not longer. Unsure if this is the way the wgan is supposed to be for Mario or Lode Runner and if not then I'm unsure why this is happening

[williamsr03]

evolve_interactive_wgan seems to now work with Mario

[williamsr03]

Also seems to work for Lode Runner as well

[williamsr03]

After testing, it seems that unconditional, conditional, and wgan all work now as attended with Lode Runner data

[williamsr03]

Since the unconditional, conditional, and wgan seem to work as attended with the Lode Runner (and Mario data), I believe this issue can be closed