Route choice in multimodal networks shows a considerable variation between different individuals as well as the current situational context. Personalization and situation awareness of recommendation algorithms are already common in many areas, e.g., online retail. However, most online routing applications still provide shortest distance or shortest traveltime routes only, neglecting individual preferences as well as the current situation. Both aspects are of particular importance in a multimodal setting as attractivity of some transportation modes such as biking crucially depends on personal characteristics and exogenous factors like the weather.As an alternative this paper introduces the FAVourite rOUte Recommendation (FAVOUR) approach to provide personalized, situation-aware route proposals based on three steps: first, at the initialization stage, the user provides limited information (home location, work place, mobility options, sociodemographics) used to select one out of a small number of initial profiles. Second, based on this information, a stated preference survey is designed in order to sharpen the profile. In this step a mass preference prior is used to encode the prior knowledge on preferences from the class identified in step one. And third, subsequently the profile is continuously updated during usage of the routing services. The last two steps use Bayesian learning techniques in order to incorporate information from all contributing individuals.The FAVOUR approach is presented in detail and tested on a small number of survey participants. The experimental results on this real-world dataset show that FAVOUR generates betterquality recommendations w.r.t. alternative learning algorithms from the literature. In particular the definition of the mass preference prior for initialization of step two is shown to provide better predictions than a number of alternatives from the literature.

No abstract

Abstract. We address the problem of automated discovery of preferred solutions by an interactive optimization procedure. The algorithm iteratively learns a utility function modeling the quality of candidate solutions and uses it to generate novel candidates for the following refinement. We focus on combinatorial utility functions made of weighted conjunctions of Boolean variables. The learning stage exploits the sparsity-inducing property of 1-norm regularization to learn a combinatorial function from the power set of all possible conjunctions up to a certain degree. The optimization stage uses a stochastic local search method to solve a weighted MAX-SAT problem. We show how the proposed approach generalizes to a large class of optimization problems dealing with satisfiability modulo theories. Experimental results demonstrate the effectiveness of the approach in focusing towards the optimal solution and its ability to recover from suboptimal initial choices.

Reactive Search Optimization advocates the adoption of learning mechanisms as an integral part of a heuristic optimization scheme. This work studies reinforcement learning methods for the online tuning of parameters in stochastic local search algorithms. In particular, the reactive tuning is obtained by learning a (near-)optimal policy in a Markov decision process where the states summarize relevant information about the recent history of the search. The learning process is performed by the Least Squares Policy Iteration (LSPI) method. The proposed framework is applied for tuning the prohibition value in the Reactive Tabu Search, the noise parameter in the Adaptive Walksat, and the smoothing probability in the Reactive Scaling and Probabilistic Smoothing (RSAPS) algorithm. The novel approach is experimentally compared with the original ad hoc reactive schemes.

No abstract