Reinforcement Learning with TensorFlow

High dimensionality problem

With the increase in the number of dimensions, data increases. As a result, there is more computation covering the complete state-action spaces.

Let's take an example:

• For each dimension, state-space is discretized in 10 different states
• Therefore, a three-dimensional state space will have 10x10x10 = 1000 states
• Thus with increase in dimensionality, the state will increase 10 fold

 

Thus, with an increase in dimensions, evaluation becomes difficult. Function approximators such as neural networks handle this problem effectively. The issues with robotic systems are high dimensional states and actions because of anthropomorphic (human-like) robots. Classical reinforcement learning approaches consider a grid-world ...