Disentangling Linear Quadratic Control with Untrusted ML Predictions

Authors: Tongxin Li, Hao Liu, Yisong Yue

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We validate the applicability of DISC across a spectrum of practical scenarios. We demonstrate the practicality of DISC through two real-world examples (Section 5): a drone navigation problem with mixed external disturbances and voltage control in a power grid with heterogeneous power injections.
Researcher Affiliation Academia Tongxin Li1 Hao Liu2 Yisong Yue2 1School of Data Science The Chinese University of Hong Kong, Shenzhen, China litongxin@cuhk.edu.cn 2Computing + Mathematical Sciences California Institute of Technology, Pasadena, USA {hliu3, yyue}@caltech.edu
Pseudocode Yes Algorithm 1: DISentangled Confidence (DISC) policy
Open Source Code Yes The code is available at https://github.com/tspbfs/Disentangle Control.
Open Datasets Yes For the regular solar and wind generation time series, we use the real solar PV generation time series data from the DTU-Data in 2021 [32] and wind generation from the U.S. Virgin Islands Wind Resources from the National Renewable Energy Lab (NREL) [33].
Dataset Splits No The paper specifies a
Hardware Specification No Our contributions focus on the theory side, and the experimental setup is sufficiently basic that it does not require intense computing resources such as GPUs.
Software Dependencies No The paper mentions using
Experiment Setup Yes Experimental Setup. To generate ML predictions, we use the Fast ICA method in [30] to decompose the mixed perturbations and train a multi-layer perceptron neural network with 4 hidden layers to predict the future latent variables. ... we implement a follow-the-regularized-leader (FTRL) optimization with a ℓ2-regularizer... The detailed parameters and hyper-parameters used in our experiments can be found in Appendix C.3. ... The warm-start buffer size is set as 100 and 50 respectively for the drone navigation and voltage control applications. ... we update the ML prediction model at each time step t [T] and use a length-b subsequence of the collected time series... to predict the future w = 5 (w is the prediction window size defined in Section 2) steps of perturbations... All prediction models are formed via a fully-connected neural network with 4 hidden layers with a width 80 and Leaky Re LU as the activation function except the final layer, and are trained using Adam [72] as the optimizer with a learning rate 1e 3 for 500 epochs. ... Table 2 below summarizes the detailed parameters used in Section C.