Since the advent of many unsupervised learning methods, as a pretraining step for the main supervised task (mostly under the name of Deep Learning), it shouldn't be strange to ask, what is the current state of "pretraining and learning from unlabeled data, for Reinforcement Learning"? Any recent/old works on this? Any suggestions for future work?
$\begingroup$
$\endgroup$
0
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
I think, you should look into the Learning from Demonstration direction. The idea is pretty simple. Let's say, we want to teach a bot to play a video game. We record a human playing the game, and give the model this data in order to pre-train it then.
There are lot's of possible ways of using this data. If you are interested in the unsupervised pre-training, you should look into the Inverse Reinforcement Learning (IRL) direction. In two words, the method tries to approximate the reward function and does the usual RL with this approximation along the road.
I'm not aware of pre-training in the IRL, but it should be possible and interesting for the investigation, from my point of view. These are some well-known works where you can start: Algorithms for IRL by Andrew Ng and Stuart Russel, and Apprenticeship Learning via IRL by Peter Abbeel and Andrew Ng.
In case you have the reward function in addition to actions, you might do a lot of interesting stuff with it, including pre-training. This work is one of the latest on the topic, and one will find much more interesting stuff in the References.
