When training an agent using the Reinforcement Learning Toolbox, how can I use a custom stopping criterion?

The current options only allow for 5 predefined choices ("AverageSteps", "AverageReward", "EpisodeReward", "GlobalStepCount", "EpisodeCount"). I want to include a stopping criterion different from these. Is there any option to do the same?

[goc3]

I was about to ask a similar question... The "accepted" answer below doesn't actually answer the question—instead, it confirms that those are the only available stop criteria.

It would be great if additional options and/or support for custom stopping criteria were added.

As an example, for a particular application, I would like to stop training once the episode reward plateaus. It is not known beforehand at what value it will plateau, so having to set a constant before training is very limiting for any application that is programmed to be dynamic or to proceed automatically based on training results.

Answer 1

[Rajani Mishra]

trainOpts = rlTrainingOptions(Name,Value) creates an option set for training using specified name-value pairs.

Arguments like - 'StopTrainingCriteria', 'StopTrainingValue', 'MaxEpisodes' should be specified for defining stopping criterion while training an agent.

• StopTrainingCriteria: Specifies the termination condition. Takes one of the choices as you have mentioned

• StopTrainingValue: Specifies the Critical value of training termination condition. Training terminates when the termination condition specified by the StopTrainingCriteria option equals or exceeds this value

• MaxEpisodes: Specifies maximum number of episodes to train the agent, once the number of episodes reached training terminates

For more information please refer to

1. https://www.mathworks.com/help/reinforcement-learning/ref/rltrainingoptions.html#mw_f2a4c660-a2a4-417d-9dbc-df113aa3b975
2. https://www.mathworks.com/help/reinforcement-learning/training-and-validation.html?s_tid=CRUX_lftnav

[Tuwe Löfström]

So there is no way of adding a custom stopping criteria, in a similar way as you can define custom reset and step functions?