Hello everyone, tonight's post isn't terribly long as for the most part it's been more of the same (AI work). Generally it's progressing well, though I do need to create a standardized set of training data which is a bit of a pain. Overall it's quite effective, and more capable than a manual simulated neural network. I even figured out a way to include testing/learning inside the game itself, so players could click "yes, makes sense" and "no, fail" and back propagate that for learning, but then realized that approach has some noticeable drawbacks (namely instance issues).
I did have another idea, however, which is to essentially do the same thing as far as player feedback on the AI goes, but instead generate a report containing the information that can then be used to train the AI all at once in a controlled fashion. This also has the advantage of being smaller/easier to send back, and it also gives me the chance to review the data for a quick "second opinion" on the AI decision.
The downside of this plan also happens to be pretty cool; namely that it depends on you to help train the AI into what you want it to be. This is pretty neat, I think, in a way your feedback lives on essentially permanently in the game AI. The downside is that the quality of the training is limited by the quality of the people creating it, so someone making random choices in regard to AI performance--or deliberately making screwy feedback--could cause more harm than good if it isn't caught.
The extra work to add this education system to the game isn't a ton, especially if it's kept on the simple side as far as the interface goes. Let me know what you think on the poll!
Here's a look at the mostly-done training setup, which also shows the main subject tags:
These are all fairly general, but the hope is that they can be sufficiently expressive when combined with the 8 modifier tags and other tags in the group. Rape, for example, would be violence, sex, and negative morality, with the NPC as the actor, probably with increased kinkiness based on the NPC's kinks. Trying to create a generalized set of tags wasn't easy or all that fun, but it seems to be workable so far.
also, since the last preview showed a somewhat more "stacked' Kazuna Ai, and I got a couple questions about that, I figured I'd show this preview of the "current status" of the art. I'm not sure if I'll be able to completely finish it by release, but I think I'll have Ai-Chan done, at least. :D
No comments:
Post a Comment