Neuro-Fuzzy Control Assignments for ECE NTUA 9th Semester Course "Neuro-Fuzzy Control and Applications" (ECE1374).
This repository contains an implementation of Q-Learning Reinforcement Learning Technique for Optimal Control Problems (Approximate Dynamic Programming). In particular, we are designing a state feedback controller for a linear dynamical system (without knowing the model) that stabilizes our system and minimizes an Quadratic Cost Criterion (LQR).
The agent tries a random input u(k) = - L x(k)
on the dynamical system x(k+1) = f(x(k), u(k))
and by sampling the response x(k)
it learns the optimal gain matrix K
for the LQR Criterion.
We use samples from the distribution and construct a least squares problem ZH =R
which is solved in order to get the unvectorized matrix H
and the gain. More information can be found in [1].
The implementation is done in MATLAB.
[1] Bradtke, Steven J. "Reinforcement learning applied to linear quadratic regulation." Advances in neural information processing systems. 1993.