Pedro Vilanova

Hybrid Chernoff Tau-Leap

Markovian pure jump processes model a wide range of phenomena, including chemical reactions at the molecular level, dynamics of wireless communication networks, and the spread of epidemic diseases in small populations. There exist algorithms such as Gillespies stochastic simulation algorithm (SSA) and Andersons modified next reaction method (MNRM) that simulate a single path with the exact distribution of the process, but this can be time consuming when many reactions take place during a short time interval. Gillespies approximated tau-leap method, on the other hand, can be used to reduce computational time, but it may lead to nonphysical values due to a positive one-step exit probability, and it also introduces a time discretization error. Here, we present a novel hybrid algorithm for simulating individual paths which adaptively switches between the SSA and the tau-leap method. The switching strategy is based on a comparison of the expected interarrival time of the SSA and an adaptive time step derive...

Mean field games for cognitive radio networks

In this paper we study mobility effect and power saving in cognitive radio networks using mean field games. We consider two types of users: primary and secondary users. When active, each secondary transmitter-receiver uses carrier sensing and is subject to long-term energy constraint. We formulate the interaction between primary user and large number of secondary users as an hierarchical mean field game. In contrast to the classical large-scale approaches based on stochastic geometry, percolation theory and large random matrices, the proposed mean field framework allows one to describe the evolution of the density distribution and the associated performance metrics using coupled partial differential equations. We provide explicit formulas and algorithmic power management for both primary and secondary users. A complete characterization of the optimal distribution of energy and probability of success is given.

Mean-field learning for satisfactory solutions

One of the fundamental challenges in distributed interactive systems is to design efficient, accurate, and fair solutions. In such systems, a satisfactory solution is an innovative approach that aims to provide all players with a satisfactory payoff anytime and anywhere. In this paper we study fully distributed learning schemes for satisfactory solutions in games with continuous action space. Considering games where the payoff function depends only on own-action and an aggregate term, we show that the complexity of learning systems can be significantly reduced, leading to the so-called mean-field learning. We provide sufficient conditions for convergence to a satisfactory solution and we give explicit convergence time bounds. Then, several acceleration techniques are used in order to improve the convergence rate. We illustrate numerically the proposed mean-field learning schemes for quality-of-service management in communication networks.

Multiscale Modeling of Wear Degradation in Cylinder Liners

Every mechanical system is naturally subjected to some kind of wear process that, at some point, will cause failure in the system if no monitoring or treatment process is applied. Since failures of...

An efficient forward–reverse expectation-maximization algorithm for statistical inference in stochastic reaction networks

ABSTRACT In this work, we present an extension of the forward–reverse representation introduced by Bayer and Schoenmakers (Annals of Applied Probability, 24(5):1994–2032, 2014) to the context of stochastic reaction networks (SRNs). We apply this stochastic representation to the computation of efficient approximations of expected values of functionals of SRN bridges, that is, SRNs conditional on their values in the extremes of given time intervals. We then employ this SRN bridge-generation technique to the statistical inference problem of approximating reaction propensities based on discretely observed data. To this end, we introduce a two-phase iterative inference method in which, during phase I, we solve a set of deterministic optimization problems where the SRNs are replaced by their reaction-rate ordinary differential equations approximation; then, during phase II, we apply the Monte Carlo version of the expectation-maximization algorithm to the phase I output. By selecting a set of overdispersed seeds as initial points in phase I, the output of parallel runs from our two-phase method is a cluster of approximate maximum likelihood estimates. Our results are supported by numerical examples.

Mean field interaction in biochemical reaction networks

In this paper we establish a relationship between chemical dynamics and mean field game dynamics. We show that chemical reaction networks can be studied using noisy mean field limits. We provide deterministic, noisy and switching mean field limits and illustrate them with numerical examples.