Parallelizing Exploration-Exploitation Tradeoffs in Gaussian Process Bandit Optimization
Authors: Thomas Desautels, Andreas Krause, Joel W. Burdick
JMLR 2014 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate GP-BUCB and GP-AUCB on several simulated and real data sets. |
| Researcher Affiliation | Academia | Thomas Desautels EMAIL Gatsby Computational Neuroscience Unit University College London Alexandra House, 17 Queen Square, London WC1N 3AR, UK Andreas Krause EMAIL Department of Computer Science ETH Zurich Universit atstrasse 6, 8092 Z urich, Switzerland Joel W. Burdick EMAIL Department of Mechanical Engineering California Institute of Technology 1200 E California Blvd., Pasadena, CA 91125, USA |
| Pseudocode | Yes | Algorithm 1 GP-UCB Algorithm 2 GP-BUCB Algorithm 3 Uncertainty Sampling Algorithm 4 GP-AUCB Algorithm 5 GP-AUCB Local |
| Open Source Code | Yes | All experiments were performed in MATLAB using custom code, which we make publicly available.4 4. See www.its.caltech.edu/~tadesaut/. |
| Open Datasets | Yes | We also test GP-BUCB and GP-AUCB on a database of Widmer et al. (2010), as considered for experimental design by Krause and Ong (2011). Lastly, we compare the algorithms on a data set of leg muscle activity triggered by therapeutic spinal electrostimulation in spinal cord injured rats. From the 3-by-9 grid of electrodes on the array, a pair of electrodes is chosen to activate, with the first element of the pair used as the cathode and the second used as the anode. |
| Dataset Splits | No | The paper describes using synthetic and real datasets, including 100 example functions drawn from a GP and data from 116 electrode pairs, but does not provide specific train/test/validation split percentages, sample counts, or methodology for data partitioning. |
| Hardware Specification | Yes | All computational time experiments were performed on a desktop computer (quad-core Intel i7, 2.8 GHz, 8 GB RAM, Ubuntu 10.04) running a single MATLAB R2012a process. |
| Software Dependencies | Yes | All computational time experiments were performed on a desktop computer (quad-core Intel i7, 2.8 GHz, 8 GB RAM, Ubuntu 10.04) running a single MATLAB R2012a process. Where applicable, the covariance function from the GPML toolbox (Ver. 3.1, Rasmussen and Nickisch, 2010) used is also listed by name. |
| Experiment Setup | Yes | in all algorithms which use the UCB or BUCB decision rules, the value of αt has been set such that it has a small premultiplier (0.05 or 0.1, see Table 2). Further, despite the rigors of analysis explored above in Section 4, we choose to set βt = αfb[t]+1 for the batch and delay algorithms, without reference to the value of C or the batch length B. Table 2: Experimental kernel functions and parameters. |