Bayesian Analysis of Combinatorial Gaussian Process Bandits

Authors: Jack Sandberg, Niklas Åkerblom, Morteza Haghir Chehreghani

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Furthermore, we employ our framework to address the challenging real-world problem of online energy-efficient navigation, where we demonstrate its effectiveness compared to the alternatives. [...] 4 EXPERIMENTS In this section, we consider the important real-world application of online energy-efficient navigation for electric vehicles and formulate it as a combinatorial and contextual bandit problem. [...] The cumulative regret is shown in Fig. 2.
Researcher Affiliation Collaboration Jack Sandberg1, Niklas Akerblom1,2 & Morteza Haghir Chehreghani1 1Department of Computer Science and Engineering, Chalmers University of Technology & University of Gothenburg 2Volvo Car Corporation EMAIL, EMAIL
Pseudocode Yes Algorithm 1 Framework for combinatorial volatile semi-bandit problem [...] Algorithm 2 Compute Rectified Indices [...] Algorithm 3 SVGP Optimization Procedure
Open Source Code No The kernels, the SVGP model and Algorithm 3 was implemented using GPy Torch (Gardner et al., 2018).
Open Datasets Yes In our experiments we use the road networks of Luxembourg and Monaco (Codeca et al., 2017; Codeca & H arri, 2018, based on data by Open Street Map contributors, 2017). Elevation data (Administration de la navigation a erienne, 2018) is added to the network using QGIS and the netconvert tool from SUMO.
Dataset Splits No The six methods are evaluated 5 times each on the four routes in the Luxembourg and Monaco networks with a horizon of T = 500. The cumulative regret is shown in Fig. 2. [...] The four parametrizations are evaluated 5 times each on the four routes in the Luxembourg and Monaco networks with a horizon of T = 500.
Hardware Specification No The computations were enabled by resources provided by the National Academic Infrastructure for Supercomputing in Sweden (NAISS), partially funded by the Swedish Research Council through grant agreement no. 2022-06725.
Software Dependencies No The kernels, the SVGP model and Algorithm 3 was implemented using GPy Torch (Gardner et al., 2018). [...] Elevation data (Administration de la navigation a erienne, 2018) is added to the network using QGIS and the netconvert tool from SUMO.
Experiment Setup Yes The graph kernel is initialized with parameters νG = 2, κG = 1 and σG set according to the prior. The natural gradient descent learning rate is set to 0.1 whilst the Adam learning rate is set to 0.01. The GP model uses a batch size B of 2500 and 1 gradient step per optimization procedure. The number of inducing points is set to 1000. [...] The noise variance ς2 is set to 0.12σ2 det for all edges and the kernel lengthscales are set to 1.