You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
One episode should consist of several trials in the same environment. The number of trials should be a parameter.
We may also want to add curriculum learning over the course of the trials, such that the agent sees the same mechanic in a more complex environment, or sees two distinct mechanics and then merges them into a single puzzle in the final trial.
This can be controlled by parameters that specify the procedural generation for each trial.
One challenge is letting the procedural generation scale up based on the current capabilities of the agent. We need environments to be challenging, but not so challenging nothing is learned. Ultimately this will probably have to be dealt with outside the environment itself, so I guess it's okay with the don't have scaling within a single environment.
The text was updated successfully, but these errors were encountered:
One episode should consist of several trials in the same environment. The number of trials should be a parameter.
We may also want to add curriculum learning over the course of the trials, such that the agent sees the same mechanic in a more complex environment, or sees two distinct mechanics and then merges them into a single puzzle in the final trial.
This can be controlled by parameters that specify the procedural generation for each trial.
One challenge is letting the procedural generation scale up based on the current capabilities of the agent. We need environments to be challenging, but not so challenging nothing is learned. Ultimately this will probably have to be dealt with outside the environment itself, so I guess it's okay with the don't have scaling within a single environment.
The text was updated successfully, but these errors were encountered: