Target-based Surrogates for Stochastic Optimization
Authors: Jonathan Wilder Lavington, Sharan Vaswani, Reza Babanezhad Harikandeh, Mark Schmidt, Nicolas Le Roux
ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | To evaluate our target optimization framework, we consider online imitation learning (OIL) as our primary example. ... Using the Mujoco benchmark suite (Todorov et al., 2012) we demonstrate that SSO results in superior empirical performance (Section 5). We then consider standard supervised learning problems where we compare SSO with different choices of the target surrogate to standard optimization methods. These empirical results indicate the practical benefits of target optimization and SSO. |
| Researcher Affiliation | Collaboration | 1University of British Columbia 2Simon Fraser University 3Samsung SAIT AI Lab, Montreal 4Canada CIFAR AI Chair (Amii) 5Microsoft Research 6 Canada CIFAR AI Chair (MILA). |
| Pseudocode | Yes | Algorithm 1 (Stochastic) Surrogate optimization |
| Open Source Code | Yes | The code is available at http://github.com/WilderLavington/Target-Based-Surrogates-For-Stochastic-Optimization. |
| Open Datasets | Yes | Using the Mujoco benchmark suite (Todorov et al., 2012) we demonstrate that SSO results in superior empirical performance (Section 5). We consider a simple supervised learning setup. In particular, we use the the rcv1 dataset from libsvm (Chang and Lin, 2011) across four different batch sizes under a logistic-loss. |
| Dataset Splits | No | The paper does not explicitly specify training, validation, and test splits with percentages or sample counts for its experiments. It mentions using datasets like rcv1 and Mujoco, which often have standard splits, but these specific split details are not provided within the paper. |
| Hardware Specification | Yes | All experiments were run using an NVIDIA Ge Force RTX 2070 graphics card. with a AMD Ryzen 9 3900 12-Core Processor. |
| Software Dependencies | No | The paper references "Paszke et al. (2019)" for default step-size, implying the use of PyTorch, but it does not specify exact version numbers for PyTorch or any other software dependencies, which is required for reproducibility. |
| Experiment Setup | Yes | For SSO, since optimization of the surrogate is a deterministic problem, we use the standard back-tracking Armijo line-search (Armijo, 1966) with the same hyper-parameters across all experiments. Each optimization algorithm is run for 500 epochs (full passes over the data). Each algorithm is evaluated over three random seeds following the same initialization scheme. |