Keras Reinforcement Learning Projects

Reinforcement learning from human preferences

Often in everyday problems, it is not possible to define a well-specified reward function. In fact, many tasks involve complex, poorly-defined, or difficult-to-specify goals. For example, suppose you want to use reinforcement learning to train a robot to clean a house. It is not easy to define an adequate reward function, which will have to depend on the data coming from the robot sensors. If we could successfully communicate our real goals to our agents, it would be a significant step toward resolving these issues.

If we have demonstrations of the desired task, we can extract a reward function using the inverse reinforcement learning. Furthermore, we can use the imitation of learning to clone ...