Constant Regret, Generalized Mixability, and Mirror Descent

Authors: Zakaria Mhammedi, Robert C. Williamson

NeurIPS 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We empirically demonstrated the performance of this algorithm on football game predictions (see Appendix J).
Researcher Affiliation Collaboration Zakaria Mhammedi Research School of Computer Science Australian National University and DATA61 zak.mhammedi@anu.edu.au Robert C. Williamson Research School of Computer Science Australian National University and DATA61 bob.williamson@anu.edu.au
Pseudocode Yes Algorithm 1: Aggregating Algorithm; Algorithm 2: Generalized Aggregating Algorithm; Algorithm 3: Adaptive Generalized Aggregating Algorithm (AGAA)
Open Source Code No The paper mentions empirical demonstration and provides an appendix for details, but does not provide concrete access to source code or explicitly state it's open-source.
Open Datasets Yes We have tested the aggregating algorithms on real data as studied by Vovk [11].
Dataset Splits No The paper mentions using real data and two datasets but does not provide specific details on training, validation, or test splits.
Hardware Specification No The paper does not provide specific hardware details (e.g., CPU/GPU models, memory) used for running experiments.
Software Dependencies No The paper does not provide specific software dependencies with version numbers.
Experiment Setup Yes Consider the Brier game G2 ℓBrier( 2, 2); a probability game with 2 experts {θ1, θ2}, 2 outcomes {0, 1}, and where the loss ℓBrier is the Brier loss [11] (which is 1-mixable). Assume that; expert θ1 consistently predicts Pr(x = 0) = 1/2; expert θ2 predicts Pr(x = 0) = 1/4 during the first 50 rounds, then switches to predicting Pr(x = 0) = 3/4 thereafter; the outcome is always x = 0. A straightforward simulation using the AGAA with the Shannon entropy, Vovk s substitution function for the Brier loss [11], βt as in Theorem 19 with vt := 1 8t Pt s=1 ℓBrier(xs, As), yields RΦ ℓBrier + Rθ (T) 5, T 150, where in this case θ = θ2 is the best expert for T 150.