Random stage selection allows more diverse sampling and exploration for agents, though the wide variety of stages in Super Mario Bros. can be challenging for agents to learn. Specifying a subset of stages to randomly sample allows the user to hand-select similar levels for agents to have more diverse observations, while still being focused enough to learn the mechanics of the environments in question.
For example, several of the castle levels have identical layouts but feature slightly different hazards. Allowing the agent to learn on different castle environments can help it be more generic while still having a reasonable chance of completing the level.
Type of change
Please select all relevant options:
[ ] Bug fix (non-breaking change which fixes an issue)
[x] New feature (non-breaking change which adds functionality)
[ ] MGMT (non-breaking change to deployment, CI, etc.
Description
Random stage selection allows more diverse sampling and exploration for agents, though the wide variety of stages in Super Mario Bros. can be challenging for agents to learn. Specifying a subset of stages to randomly sample allows the user to hand-select similar levels for agents to have more diverse observations, while still being focused enough to learn the mechanics of the environments in question.
For example, several of the castle levels have identical layouts but feature slightly different hazards. Allowing the agent to learn on different castle environments can help it be more generic while still having a reasonable chance of completing the level.
Type of change
Please select all relevant options:
Checklist