#Changing max step (max episode step) dynamically per lesson using curriculum learning

1 messages · Page 1 of 1 (latest)

snow scroll
#

I'm using curriculum learning with PPO to tackle obstacle avoidance. I have usually set the max episode this step value to a fixed number which for initial lessons i find quite large, in order to time out the agent wanders aimlessly for a while, this ain't particularly useful to me nor the training. Now my question is: is there a way and is it advisable to set the max steps of the academy to a lower intital amount and then gradually increase them as the curriculum progresses? And if so, would there be adverse effects for the agent's learning if this value is not observable? Thank you 🙂