A Composite Randomized Incremental Gradient Method

Authors: Junyu Zhang, Lin Xiao

ICML 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We provide numerical experiments in Section 5. In this part, we present numerical experiments for two applications: risk-averse portfolio optimization and policy evaluation for Markov decision processes. Figure 1. Experiments on the risk-averse portfolio optimization problem. Figure 2. Experiments on MDP policy evaluation problem.
Researcher Affiliation Collaboration 1Department of Industrial and Systems Engineering, University of Minnesota, Minneapolis, Minnesota, USA. 2Microsoft Research, Redmond, Washington, USA.
Pseudocode Yes Algorithm 1 Composite SAGA (C-SAGA)
Open Source Code No The paper does not include an unambiguous statement that the authors are releasing the source code for the methodology described, nor does it provide a direct link to a code repository.
Open Datasets No In our experiments, the reward vectors Ri are first generated as n i.i.d Gaussian random vectors with a random correlation matrix C = LLT, where L Rd d satisfies N(0, 1) distribution elementwise. In the experiments, Pπ, Φ and Rπ are generated randomly.
Dataset Splits No The paper uses randomly generated data for its experiments and does not provide specific dataset split information (percentages, sample counts, or predefined splits) for training, validation, or testing.
Hardware Specification No The paper does not provide specific hardware details (e.g., exact GPU/CPU models, processor types, or memory amounts) used for running its experiments.
Software Dependencies No The paper does not provide specific ancillary software details, such as library names with version numbers, needed to replicate the experiment.
Experiment Setup Yes Both VRSC-PG and C-SAGA use the same step size η = 0.001 and batch size s = n2/3 . They are chosen from by experimenting with {1, 0.1, 0.01, 0.001, 0.0001}, and η = 0.001 works best for VRSC-PG and for C-SAGA. For the case where S = 10, both VRSC-PG and C-SAGA use the same batch size s = 1. C-SAGA takes a step size η = 0.1, while VRSC-PG takes a stepsize of η = 0.03, because it diverges under η = 0.1 and η = 0.03 seems to work best VRSC-PG. For S = 100, we set η = 0.005 and batch size s = 10 for C-SAGA and VRSC-PG. The step size is chosen as the best among {0.1, 0.05, 0.01, 0.005, 0.001}.