Scalable Identification of Partially Observed Systems with Certainty-Equivalent EM
Authors: Kunal Menda, Jean De Becdelievre, Jayesh Gupta, Ilan Kroo, Mykel Kochenderfer, Zachary Manchester
ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We empirically show that the certainty-equivalent approximation to expectation-maximization can be a reliable and scalable approach for highdimensional deterministic systems, which are common in robotics. We formulate certaintyequivalent expectation-maximization as block coordinate-ascent, and provide an efficient implementation. The algorithm is tested on a simulated system of coupled Lorenz attractors, demonstrating its ability to identify high-dimensional systems that can be intractable for particle-based approaches. Our approach is also used to identify the dynamics of an aerobatic helicopter. |
| Researcher Affiliation | Academia | 1Stanford University, CA, USA. |
| Pseudocode | Yes | Algorithm 1. CE-EM Implementation |
| Open Source Code | Yes | The codebase for this work is available at https://github.com/sisl/CEEM. A codebase implementing CE-EM and other supplementary material can be found at our website: https://sites.google.com/stanford.edu/ceem/. |
| Open Datasets | Yes | We use data collected by the Stanford Autonomous Helicopter Project (Abbeel et al., 2010). Menda, K., de Becdeli evre, J., Gupta, J. K., Kroo, I., Kochenderfer, M. J., and Manchester, Z. Normalized Stanford helicopter dataset, February 2020. URL https://doi.org/10.5281/zenodo.3662987. |
| Dataset Splits | Yes | Trajectories are split into 10 s long chunks and then randomly distributed into train, test, and validation sets according to the established protocol (Abbeel et al., 2010; Punjani & Abbeel, 2015) and summarized in Appendix A.1. The train, test, and validation sets respectively contain 466, 100, and 101 trajectories of 500 time-steps each. |
| Hardware Specification | Yes | Experiments were run on a computer with an Intel R Core TM i7-6700K CPU @ 4.00GHz 8 processor and 32GB of memory. |
| Software Dependencies | No | The paper mentions 'MATLAB command n4sid' and 'SciPy: Open source scientific tools for Python, 2001 . URL http: //www.scipy.org/', but does not provide specific version numbers for these or other software dependencies. |
| Experiment Setup | Yes | Np = 100 particles for filtering, and sample Ns = 10 smoothed trajectories using FFBSi. We randomly initialize each element of the parameters being optimized (θ = [σ1:K, ρ1:K, β1:K, H]) to within 10% of the their value in θtrue. The values of σw and σv are treated as hyperparmeters of CE-EM, and are both set to 1. |