search

nazaruka / gym-http-api

NSGA2-based Sonic agent + experimental code
MIT License
1 stars 1 forks source link

Behavior archive #38

Closed schrum2 closed 5 years ago

schrum2 commented 5 years ago

Been test running Lamarckian evolution. I notice that although progress is made (slowly) there is a problem in NSGA-II's selection: specifically, really crappy points look very "novel" as the search goes on. We need an option (command line parameter) that keeps all behavior characterizations in an archive (like with Novelty Search) that is used for diversity/novelty calculations. This should make those really poor points not selected for.

schrum2 commented 5 years ago

Something weird is going on with this. The behavior characterization for a single agent has 4 components, which means that it is somehow merging characterizations when an agent gets evaluated in the second generation. Need to diagnose.

It's probably also a good idea to log the whole behavior archive every generation as well. Not only will this be good for troubleshooting, it would be required for resuming an aborted run.

schrum2 commented 5 years ago

I'm saving the archive now.

Regarding that weird situation where one behavior vector was longer than the others ... this may have been a freak occurrence, which means it will be even harder to track it down. I haven't seen it again, but maybe saving the archive will help track this down.

schrum2 commented 5 years ago

Extensive work on this. I think it is complete, but needs to be tested.