allow ending the episode for MaxStepsReached

[chriselion]

Proposed change(s)

This came up when trying to use MaxSteps on the match-3 sample. If you have multiple Agents in the scene using on-demand decisions, their step count will get incremented whenever the Academy steps, even if they're not doing anything. For training, you can work around this by requesting a decision each step, but for inference or heuristic mode when you might want to show animations of the Agent "boards", the decision request frames get out of sync and you end up hitting MaxSteps too soon.

I think this is the best solution without breaking existing behavior (a better approach would be to only count decision or action steps towards MaxSteps, but that's a breaking change). An alternative would be to make DoneReason public and allow that as an optional parameter to EndEpisode.

Useful links (Github issues, JIRA tickets, ML-Agents forum threads etc.)

https://jira.unity3d.com/browse/MLA-1345

Types of change(s)

• New feature

Checklist

• Added tests that prove my fix is effective or that my feature works
• Updated the changelog (if applicable)
• Updated the documentation (if applicable)
• Updated the migration guide (if applicable)

Other comments

[vincentpierre]

I think the logic is cleaner, I have some reservations on adding the new public method. Happy to hear you thoughts.

[vincentpierre]

I would prefer not calling this MaxStepReached, but call it Interrupted. The reason is that MaxStepReached will not (I hope) be the only reason an agent will be interrupted forever. A user triggered scene reset should have the same effect as MaxStepReached (for example if the Agents reset but not because it is their fault)

[chriselion]

ok, so just rename this to EpisodeInterrupted (and change the docs accordingly)?