Efficient Parallelized Simulation of Cyber-Physical Systems

Authors: Bas van der Heijden, Laura Ferranti, Jens Kober, Robert Babuska

TMLR 2024 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We validate our approach on two real-world systems and perform an extensive ablation, demonstrating superior performance compared to baseline methods. An ablation study on the effects of the algorithmic simplifications was also conducted and included as an appendix.
Researcher Affiliation Academia Bas van der Heijden EMAIL Cognitive Robotics Delft University of Technology; Laura Ferranti EMAIL Cognitive Robotics Delft University of Technology; Jens Kober EMAIL Cognitive Robotics Delft University of Technology; Robert Babuska EMAIL Cognitive Robotics, Delft University of Technology, CIIRC, Czech Technical University in Prague
Pseudocode Yes Algorithm 1: Minimum Common Supergraph Search (mcs) Algorithm 2: Largest Match Search
Open Source Code Yes Finally, a motivational video for our approach is included as supplementary material, together with example source code to efficiently compile parallelized training environments with our approach.
Open Datasets No The paper describes using a "pendulum swing-up task" which is a "well-known RL benchmark", and for other experiments, it states "We record 10 computation graphs from the real-world system" and "We artificially generate computation graphs". No explicit links, DOIs, repositories, or formal citations are provided for accessing the datasets themselves that were used or created for the experiments.
Dataset Splits No The paper mentions recording or artificially generating computation graphs for experiments (e.g., "We record 10 computation graphs from the real-world system", "generate 10 computation graphs, each running for a duration of 10 seconds"). However, it does not specify how these graphs or any other data were split into training, validation, or test sets for reproducibility.
Hardware Specification Yes We carried out a parallelized performance evaluation of the swing up-task on an RTX 3070 GPU. We implement our approach using JAX (Frostig et al., 2018) and execute it on an RTX 3070 GPU. We employ Alg. 1 to identify a supergraph on a single core of an Intel Core i9-10980HK
Software Dependencies No The paper mentions several software components like JAX, soft actor-critic (SAC), Cross Entropy Method (CEM), and Brax, all with citations. However, specific version numbers for these software components are not provided (e.g., "JAX (Frostig et al., 2018)" does not specify the JAX version).
Experiment Setup Yes Policies were trained using soft actor-critic (SAC) (Haarnoja et al., 2018) in two simulators: one emulating delays (our approach: mcs) and another without delays (sequential approach: seq). Our implementation employs CEM for three iterations, involving 75 samples per iteration and a planning horizon of two control steps, each lasting 0.15 seconds.