Idea: We can turn every procgen experiment into a memory experiment by repeating it. For this we evaluate the same individual n times without resetting its neural state between the evaluation runs. During the evolution we ignore the fitness score on the first n-1 runs and only record the fitness from the n-th run.
todo:
[ ] come up with suitable vocabulary for the concepts in this issue
[ ] add configuration parameter to episode runner. something like number_memory_runs_before_actual_evaluation: int = 0
[ ] consume this parameter in the episode-runner accordingly
The experiment is a success if individuals perform better on the n-th run then on their first run.
Ideas for follow-up features:
record the difference in fitness between the first and the n-th run for the training-curve. I suspect, that the benefit from the memory-runs only kicks in later generation.
Idea: We can turn every procgen experiment into a memory experiment by repeating it. For this we evaluate the same individual n times without resetting its neural state between the evaluation runs. During the evolution we ignore the fitness score on the first n-1 runs and only record the fitness from the n-th run.
todo:
number_memory_runs_before_actual_evaluation: int = 0
The experiment is a success if individuals perform better on the n-th run then on their first run.
Ideas for follow-up features:
maybe related issue #39