Model-free and Bayesian Ensembling Model-based Deep Reinforcement Learning for Particle Accelerator Control Demonstrated on the FERMI FEL

Contact: simon.hirlaender(at)sbg.ac.at

Pre-print https://arxiv.org/abs/2012.09737

Please cite code as:

The included scripts:

1. To run the NAF2 as used in the paper on the pendulum run: run_naf2.py
2. To run the AE-DYNA as used in the paper on the pendulum run: AEDYNA.py
3. To run the AE-DYNA with tensorflow 2 on the pendulum run: AE_Dyna_Tensorflow_2.py

The rest should be straight forward, otherwise contact us.

These are the results of RL tests @FERMI-FEL

The problem has four degrees of freedom in state and action space. A schematic overview:

Algorithm	Type	Representational power	Noise resistive	Sample efficiency
NAF	Model-free	Low	No	High
NAF2	Model-free	Low	Yes	High
ME-TRPO	Model-based	High	No	High
AE-DYNA	Model-based	High	Yes	High

Experiments done on the machine:

A new implementation of the NAF with double Q learning (single network dashed, double network solid):

A new implementation of a AE-DYNA:

A variant of the ME-TRPO:

The evolution as presented at GSI Towards Artificial Intelligence in Accelerator Operation:

Experiments done on the inverted pendulum openai gym environment:

Cumulative reward of different NAF implementations on the inverted pendulum with artificial noise.

Comparison of the inclusion of aleatoric noise in the AE-DYNA in the noisy inverted pendulum:

Comparison of the inclusion of aleatoric noise in the AE-DYNA in the noisy inverted pendulum:

Sample efficiency of NAF and AE-DYNA:

Free run on the inverted pendulum:

Update of AE-Dyna-(SAC) to Tensorflow 2

Finally, there is an update of the AE-dyna to use tensorflow 2. Run the script AE_Dyna_Tensorflow_2.py. It is based on tensor_layers tensorlayer, which has to be installed. The script AE_Dyna_Tensorflow_2.py runs on the inverted pendulum and produces results like shown in the figure below.

If you have questions do not hesitate to contact us.