Reinforcement Learning: How to use neural network toolbox for action value approximation?

1 view (last 30 days)
I need to use action value approximation Q-learning with a neural network. I have a little knowledge of Reinforcement Learning (RL) but I have never used neural networks and I would like to use only built-in MATLAB commands. From literature research, I understand that I have to update the network weights incrementally each time I take an action, reach a state and get a reward. My state is a two dimensional vector with entries characterizing 15 possible degradation values. I have 25 possible actions, but some of them cannot be performed in some states. I think I should use a network with two inputs and 25 outputs. How do I initialize the network weights properly? And then, how can I update the weight values during RL simulation? I have used adapt and train with 'trainr' option with no success. Thank you

Answers (0)

Categories

Find more on Sequence and Numeric Feature Data Workflows in Help Center and File Exchange

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!