Reinforcement Learning with TensorFlow

Real-world challenges

Robots interact with the real physical world. Thus, a genuine problem with robot reinforcement learning is to deal with these real-world problems. This is because of regular wear and tear in the real world of robot components, which are expensive. The continuous maintenance and repair comes at a great cost in terms of labor and loss of time in maintenance and repair. Thus, safe exploration is the key issue during the learning process in robot reinforcement learning.

Perkins and Barto (2002) came up with a method for constructing reinforcement learning agents based on Lyapunov functions (Appendix A, Further topics in Reinforcement Learning). The challenges posed by the real world include changes of environmental factors, ...