Online Learning for Load Balancing of Unknown Monotone Resource Allocation Games

Authors: Ilai Bistritz, Nicholas Bambos

ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we provide numerical simulations that include the two applications discussed in Section 5. In the first two experiments, to simulate a vanishing estimation error (Definition 3), at iteration t, Gaussian noise with mean δt = 0.25 (t+1)0.4 and variance σ2 = 1 4 was added to the gradients of each player. To quantify the effectiveness of our algorithm, we compare it to the uncontrolled system where αk = 0 for all k. For all experiments, we used the step-size sequence ηt = η0 (t+1)p and the control step-size sequence εt = ε0 (t+1)q with different values of η0, ε0. We ran 100 realizations for each experiment and plotted the average result along with the standard deviation region, which was always small.
Researcher Affiliation Academia 1Department of Electrical Engineering, Stanford University. Correspondence to: Ilai Bistritz <bistritz@stanford.edu>.
Pseudocode Yes Algorithm 1 Online Load Balancing with Bandit Feedback Initialization: Let x0 X and α0 RK + . Let {ηt} , {εt} satisfy the conditions of Theorem 1. Input: Target total load vector l . For each turn t 1 do 1. Each player n updates its action using gn,t 1 and αt 1 to approximate xnun (x; α): xn,t = ΠXn xn,t 1 + ηt 1 gn,t 1 αt 1 (5) where ΠXn is the Euclidean projection into Xn. 2. The manager observes PN n=1 xk n,t for each k. 3. The manager updates the pricing coefficients using αt 1 + εt 1 PN n=1 xn,t l !#+ where [x]+ = max {x, 0} (element-wise).
Open Source Code No The paper does not provide any explicit statements about releasing source code for the methodology, nor does it include links to a code repository.
Open Datasets No The paper describes setting up simulations with randomly generated parameters and conditions (e.g., “The location y1 n of transmitter n was chosen uniformly at random on a 2D square of area 2N.”), rather than utilizing or providing access information for a publicly available dataset.
Dataset Splits No The paper performs numerical simulations and refers to “100 realizations” but does not specify any training, validation, or test dataset splits or percentages.
Hardware Specification No The paper does not provide any specific details about the hardware (e.g., GPU/CPU models, memory) used to run the simulations or experiments.
Software Dependencies No The paper does not provide specific software dependencies with version numbers, such as programming languages, libraries, or frameworks used for implementation.
Experiment Setup Yes For all experiments, we used the step-size sequence ηt = η0 (t+1)p and the control step-size sequence εt = ε0 (t+1)q with different values of η0, ε0.