| Cite the paper describing the algorithm as |
|---|
| Varga, B, Kulcsár, B, Chehreghani, MH. Deep Q-learning: A robust control approach. Int J Robust Nonlinear Control. 2022; 1- 19. doi:10.1002/rnc.6457 |
| Links |
| https://onlinelibrary.wiley.com/doi/10.1002/rnc.6457 |
| https://arxiv.org/abs/2201.08610 |
Description: Source code for the controlled DQN algorithm in the Cartpole and Acrobot environments. The training is implemented in Python using Torch and OpenAI Gym environments. The DQN_agent.py class contains all three controllers (H2, Hinf, fixed structure Hinf). The scripts Controlled_Cartpole.py and Controlled_Acrobot.py implement the training in the respective environments. Which controlled algorithm to use is selected via:
0: H2
1: Fixed Hinf
2: Dynamic Hinf
CONTROLLER = 1
The Cartpole_Plotter.py script implements some advanced plotting functions.
The control design was done with Matlab's Robust Control Toolbox for Dynamic Hinf and Fixed Hinf (code in Robust_Hinf_dynamical.m Robust_Hinf_fixedstruct.m). Controller gains were copied manually to DQN_agent.py.
