Reinforcement Learning with TensorFlow

Issues due to model uncertainty

In order to avoid the cost associated with real-world interactions, simulators are used. The catch is that the simulating model should be close to a real-world scenario. For an ideal setting, the approach is to perform the learning tasks in the simulation and transfer the knowledge model to the robot. Creating a good accurate learning model for the robot and the simulating environment model of the real-world scenario is highly challenging as it requires a huge amount of real-world data samples.

Small models learned on small sets of data leads to under-modeling, causing the robot to diverge easily from the real-world system. The issue with the simulators is that they can't replicate the real-world complexities ...