Keep a record of the observed delta and maximize reward for it.
Also add Noop actions which is turned off by default.
Let finish the execution also when no solution is found, as we will take the
minimum observed delta as result.
This is done on purpose to avoid guessing "when" is a good time to stop the agent,
as it could be in the middle of picking up a new action which is not the final
(but we need limits, we can't let it run forever).
