Constructive Preference Elicitation Over Hybrid Combinatorial Spaces

Authors: Paolo Dragone, Stefano Teso, Andrea Passerini

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

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, we demonstrate its effectiveness by empirical evaluation against existing competitors on constructive scenarios of increasing complexity.
Researcher Affiliation Collaboration Paolo Dragone University of Trento, Italy TIM-SKIL, Trento, Italy paolo.dragone@unitn.it Stefano Teso KU Leuven, Belgium stefano.teso@cs.kuleuven.be Andrea Passerini University of Trento, Italy andrea.passerini@unitn.it
Pseudocode Yes Algorithm 1 The Choice Perceptron (CP) algorithm.
Open Source Code Yes The complete experimental setting can be retrieved from: https://github.com/unitn-sml/choice-perceptron
Open Datasets Yes We evaluated all methods on the synthetic constructive benchmark introduced in (Teso, Passerini, and Viappiani 2016). ... In the second experiment, we compared CP and SETMARGIN on a much larger recommendation task, also from (Teso, Passerini, and Viappiani 2016). ... Finally, we evaluated CP on a slightly modified version of the touristic trip planning task introduced in (Teso, Dragone, and Passerini 2017).
Dataset Splits Yes In practice we also employ an adaptive Perceptron step size, which is adapted at each iteration t >= 3 from the set {0.1, 0.2, 0.5, 1, 2, 5, 10} via cross-validation on the collected feedback; it was found to work well empirically.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory) used for running the experiments.
Software Dependencies No The paper mentions 'Gecode via its Mini Zinc interface' but does not specify version numbers for these software components, which is required for reproducibility.
Experiment Setup Yes As argued in the previous section, for CP we set γ to 1/t in all experiments, in order to allow more exploration earlier on during the search. ... In practice we also employ an adaptive Perceptron step size, which is adapted at each iteration t >= 3 from the set {0.1, 0.2, 0.5, 1, 2, 5, 10} via cross-validation on the collected feedback... We set λ = 1 as in (Teso, Passerini, and Viappiani 2016). ... To help keeping running times low, the query selection procedure of CP is executed with a 20 seconds time cutoff.