Ml-agents: No episode was completed since last summary

[jek343]

Hi!

I'm trying to train a simple game modeled off of Atari pong. The first time I tried to train it, after 12000 steps it keeps telling me "no episode was completed since last summary." I tried to train it again and now every single time it tells me the same "no episode was completed since last summary." I don't know what would make there be no completed episodes especially since the first time I ran the training it worked for the first 12000 steps. Any ideas?

Here is a screenshot from the first training:

Here is a screenshot from the second training:

[mmattar]

Hi @jek343 - a couple of clarification questions: 1. Are you using On Demand Decision making? 2. What was max steps set to in the Academy?

[TejaswiniiB]

Hi, I too have this problem.. How to properly end an episode without setting Max Steps in academy ? I tried doing agent.done() for all agents and doing academy.done(), but still getting as "No episonde was completed since last summary.". I am not using onDemandDecision making.

[xiaomaogy]

@jek343 @TejaswiniiB Usually this happens when your Done() function never gets triggered. You can check if that's the case within your game.

[Dongseon-SHIN]

@xiaomaogy
Hi xiaomaogy. I have a question about "Step"!
Is 1 step means that Done() function is triggerd once?
If it is right, is there max time(maybe frame in unity) that steps finished with "No episode was completed since last summary"?

[PureRockets]

Sorry for the delay, are you still receiving this error? Can you confirm that Done() is indeed being called, perhaps place a Debug.Log statement just prior? You can see an example in Ball3DAgents.cs, I get the same error if I comment out Done()

[xiaomaogy]

@Dongseon-SHIN 1 step means that CollectObservation() function is triggered once, and that observation is sent to python, python uses the observation to generate an action and send back to C#, and AgentAction() function gets called using that action from python. Done() function only gets called when you want to do reset().

[xiaomaogy]

Also the cause for this "No episonde was completed since last summary" is due to Done() never gets called, because Done() is like: "ok, my current episode ends, now you can start using the data from this episode to calculate the cumulative reward". If Done() is never called, cumulative reward will never gets calculated, and data from this episode will never be used for training.

[xiaomaogy]

Thanks for reaching out to us. Hopefully you were able to resolve your issue. We are closing this due to inactivity, but if you need additional assistance, feel free to reopen the issue.

[TashaSkyUp]

I get the same error. Done() is most definitely being called... however restarting the training with --load solved it.

[TheDozMan]

Also, if you have more than one Brain ( ie Player brain and a Learning Brain) , the agent might be connected to the wrong one

[lock[bot]]

This thread has been automatically locked since there has not been any recent activity after it was closed. Please open a new issue for related bugs.