Custom Summaries

Environment-specific info

It is often useful to monitor custom training metrics, i.e. certain environment-specific aspects of agent's performance.

You can add custom monitored metrics by adding info["episode_extra_stats"] = { ... } to the environment's info dictionary returned from the step() function on the last step of the episode.

See sf_examples/dmlab/wrappers/reward_shaping.py for example. Here we add information about agent's performance on individual levels in DMLab-30.

Custom metrics

You can add completely custom metrics that are calculated based on other metrics or the RL algorithm state. To do this, add a custom algo observer that overrides extra_summaries() function.

See sf_examples/dmlab/train_dmlab.py where we define DmlabExtraSummariesObserver that aggregates custom environment metrics to produce a single "Human-normalized score" summary.