Real-Time Optimisation for Online Learning in Auctions

Authors: Lorenzo Croissant, Marc Abeille, Clement Calauzenes

ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We study performance for stationary bidders in Sec. 4 with 1/t convergence rate to the monopoly price, and for nonstationary bidders in Sec. 5 with O(T) dynamic regret. ... On Fig. 1, we illustrate the effect of this smoothing on p, ΠˆFt, and ΠF. ... Figure 2. Stationary case. Numerical behaviour of V-CONVOGA for different σt on i.i.d samples from a Kumaraswamy (1, 0.4). Top: averaged convergence speeds of instant regret (log-log scale). Bottom: representative reserve price trajectories. ... Figure 3. Non-stationary case. Tracking by CONV-OGA of three Kumaraswamy distributions (with parameters (1, 4), (1, 0.4), and (1, 1) resp.) with different Gaussian kernels and learning rates.
Researcher Affiliation Industry Lorenzo Croissant 1 Marc Abeille 1 Cl ement Calauz enes 1 1Criteo AI Lab, Paris, France. Correspondence to: Lorenzo Croissant <ld.croissant@criteo.com>.
Pseudocode Yes Algorithm 1 CONV-OGA input: r0, {γt}t N, k K, C [0, b] for t = 1 to + do observe bt rt proj C (rt 1 + γt pk(rt 1, bt)) ... Algorithm 2 V-CONV-OGA input: r0, {γt}t N, {kt}t N KN, C [0, b] for t = 1 to + do observe bt rt proj C (rt 1 + γt pkt(rt 1, bt))
Open Source Code No No explicit statement or link indicating the availability of the authors' source code was found.
Open Datasets No The paper mentions generating samples from a Kumaraswamy distribution for numerical behavior analysis, but does not refer to a publicly available dataset with concrete access information (link, DOI, citation).
Dataset Splits No The paper does not explicitly describe training, validation, and test dataset splits with specific percentages or counts.
Hardware Specification No No specific hardware details (e.g., GPU/CPU models, memory) used for running experiments were mentioned.
Software Dependencies No No specific software dependencies with version numbers were provided.
Experiment Setup Yes Figure 2. Stationary case. Numerical behaviour of V-CONVOGA for different σt on i.i.d samples from a Kumaraswamy (1, 0.4). Top: averaged convergence speeds of instant regret (log-log scale). Bottom: representative reserve price trajectories. ... To optimise the stationary regime we face a bias-variance trade-off. ... For zero-mean Gaussian kernels, we have Cor. 1. If we fix γt 1/t, and let {kt}t N be Gaussian (0, 1/t) kernels in Thm. 2 we have for all t 2 that: E rt r 2 = O t 1/2 . ... Figure 3. Non-stationary case. Tracking by CONV-OGA of three Kumaraswamy distributions (with parameters (1, 4), (1, 0.4), and (1, 1) resp.) with different Gaussian kernels and learning rates. ... Increasing γ shortens it but increases the width of the band of the asymptotic regime as C(γ, k) increases with γ (blue vs. green curves). For a fixed γ, the stationary regime in terms of k exhibits a bias-variance trade-off: K 1R+ 1 corresponds to the bias and k to the variance (see Prop. 3). In the case of a Gaussian kernel, increasing σ reduces variance but increases bias (green vs. red curve).