arXiv subjects

Patrick Jaillet

Publications and source records attributed to Patrick Jaillet.

At least 109 records · Page 6Linked to original sources

arXiv2016

No-Regret Learnability for Piecewise Linear Losses

In the convex optimization approach to online regret minimization, many methods have been developed to guarantee a $O(\sqrt{T})$ bound on regret for subdifferentiable convex loss functions with bounded subgradients, by using a reduction to linear loss functions. This suggests that linear loss functions tend to be the hardest ones to learn against, regardless of the underlying decision spaces. We investigate this question in a systematic fashion looking at the interplay between the set of possible moves for both the decision maker and the adversarial environment. This allows us to highlight sharp distinctive behaviors about the learnability of piecewise linear loss functions. On the one hand, when the decision set of the decision maker is a polyhedron, we establish $Ω(\sqrt{T})$ lower bounds on regret for a large class of piecewise linear loss functions with important applications in online linear optimization, repeated zero-sum Stackelberg games, online prediction with side information, and online two-stage optimization. On the other hand, we exhibit $o(\sqrt{T})$ learning rates, achieved by the Follow-The-Leader algorithm, in online linear optimization when the boundary of the decision maker's decision set is curved and when $0$ does not lie in the convex hull of the environment's decision set. Hence, the curvature of the decision maker's decision set is a determining factor for the optimal learning rate. These results hold in a completely adversarial setting.

cs.LG↗

arXiv2016

Solving Combinatorial Games using Products, Projections and Lexicographically Optimal Bases

In order to find Nash-equilibria for two-player zero-sum games where each player plays combinatorial objects like spanning trees, matchings etc, we consider two online learning algorithms: the online mirror descent (OMD) algorithm and the multiplicative weights update (MWU) algorithm. The OMD algorithm requires the computation of a certain Bregman projection, that has closed form solutions for simple convex sets like the Euclidean ball or the simplex. However, for general polyhedra one often needs to exploit the general machinery of convex optimization. We give a novel primal-style algorithm for computing Bregman projections on the base polytopes of polymatroids. Next, in the case of the MWU algorithm, although it scales logarithmically in the number of pure strategies or experts $N$ in terms of regret, the algorithm takes time polynomial in $N$; this especially becomes a problem when learning combinatorial objects. We give a general recipe to simulate the multiplicative weights update algorithm in time polynomial in their natural dimension. This is useful whenever there exists a polynomial time generalized counting oracle (even if approximate) over these objects. Finally, using the combinatorial structure of symmetric Nash-equilibria (SNE) when both players play bases of matroids, we show that these can be found with a single projection or convex minimization (without using online learning).

cs.LG↗

arXiv2016

Robust Adaptive Routing Under Uncertainty

We consider the problem of finding an optimal history-dependent routing strategy on a directed graph weighted by stochastic arc costs when the objective is to minimize the risk of spending more than a prescribed budget. To help mitigate the impact of the lack of information on the arc cost probability distributions, we introduce a robust counterpart where the distributions are only known through confidence intervals on some statistics such as the mean, the mean absolute deviation, and any quantile. Leveraging recent results in distributionally robust optimization, we develop a general-purpose algorithm to compute an approximate optimal strategy. To illustrate the benefits of the robust approach, we run numerical experiments with field data from the Singapore road network.

cs.DS↗

arXiv2015

Gaussian Process Planning with Lipschitz Continuous Reward Functions: Towards Unifying Bayesian Optimization, Active Learning, and Beyond

This paper presents a novel nonmyopic adaptive Gaussian process planning (GPP) framework endowed with a general class of Lipschitz continuous reward functions that can unify some active learning/sensing and Bayesian optimization criteria and offer practitioners some flexibility to specify their desired choices for defining new tasks/problems. In particular, it utilizes a principled Bayesian sequential decision problem framework for jointly and naturally optimizing the exploration-exploitation trade-off. In general, the resulting induced GPP policy cannot be derived exactly due to an uncountable set of candidate observations. A key contribution of our work here thus lies in exploiting the Lipschitz continuity of the reward functions to solve for a nonmyopic adaptive epsilon-optimal GPP (epsilon-GPP) policy. To plan in real time, we further propose an asymptotically optimal, branch-and-bound anytime variant of epsilon-GPP with performance guarantee. We empirically demonstrate the effectiveness of our epsilon-GPP policy and its anytime variant in Bayesian optimization and an energy harvesting task.

stat.ML↗

arXiv2015

Container Relocation Problem: Approximation, Asymptotic, and Incomplete Information

The Container Relocation Problem (CRP) is concerned with finding a sequence of moves of containers that minimizes the number of relocations needed to retrieve all containers respecting a given order of retrieval. While the problem is known to be NP-hard, certain algorithms such as the A* search and heuristics perform reasonably well on many instances of the problem. In this paper, we first focus on the A* search algorithm, and analyze lower and upper bounds that are easy to compute and can be used to prune nodes. Our analysis sheds light on which bounds result in fast computation within a given approximation gap. We present extensive simulation results that improve upon our theoretical analysis, and further show that our method finds the optimum solution on most instances of medium-size bays. On "hard" instances, our method finds an approximate solution with a small gap and within a time frame that is fast for practical applications. We also study the average-case asymptotic behavior of the CRP where the number of columns grows. We calculate the expected number of relocations in the limit, and show that the optimum number of relocations converges to a simple and intuitive lower-bound. We further study the CRP with incomplete information by relaxing the assumption that the order of retrieval of all containers are initially known. This assumption is particularly unrealistic in ports without an appointment system. We assume that the retrieval order of a subset of containers is known initially and the retrieval order of the remaining containers is observed later at a given specific time. Before this time, we assume a probabilistic distribution on the retrieval order of unknown containers. We combine the A* algorithm with sampling technique to solve this two-stage stochastic optimization problem. We show that our algorithm is fast and the error due to sampling and pruning is reasonably small.

cs.DS↗

arXiv2015

Managing Relocation and Delay in Container Terminals with Flexible Service Policies

We introduce a new model and mathematical formulation for planning crane moves in the storage yard of container terminals. Our objective is to develop a tool that captures customer centric elements, especially service time, and helps operators to manage costly relocation moves. Our model incorporates several practical details and provides port operators with expanded capabilities including planning repositioning moves in off-peak hours, controlling wait times of each customer as well as total service time, optimizing the number of relocations and wait time jointly, and optimizing simultaneously the container stacking and retrieval process. We also study a class of flexible service policies which allow for out-of-order retrieval. We show that under such flexible policies, we can decrease the number of relocations and retrieval delays without creating inequities.

cs.DS↗

arXiv2014

Distributed Multi-Depot Routing without Communications

We consider and formulate a class of distributed multi-depot routing problems, where servers are to visit a set of requests, with the aim of minimizing the total distance travelled by all servers. These problems fall into two categories: distributed offline routing problems where all the requests that need to be visited are known from the start; distributed online routing problems where the requests come to be known incrementally. A critical and novel feature of our formulations is that communications are not allowed among the servers, hence posing an interesting and challenging question: what performance can be achieved in comparison to the best possible solution obtained from an omniscience planner with perfect communication capabilities? The worst-case (over all possible request-set instances) performance metrics are given by the approximation ratio (offline case) and the competitive ratio (online case). Our first result indicates that the online and offline problems are effectively equivalent: for the same request-set instance, the approximation ratio and the competitive ratio differ by at most an additive factor of 2, irrespective of the release dates in the online case. Therefore, we can restrict our attention to the offline problem. For the offline problem, we show that the approximation ratio given by the Voronoi partition is m (the number of servers). For two classes of depot configurations, when the depots form a line and when the ratios between the distances of pairs of depots are upper bounded by a sublinear function f(m) (i.e., f(m) = o(m)), we give partition schemes with sublinear approximation ratios O(log m) and Θ(f(m)) respectively. We also discuss several interesting open problems in our formulations: in particular, how our initial results (on the two deliberately chosen classes of depots) shape our conjecture on the open problems.

cs.DC↗

arXiv2014

Parallel Gaussian Process Regression for Big Data: Low-Rank Representation Meets Markov Approximation

The expressive power of a Gaussian process (GP) model comes at a cost of poor scalability in the data size. To improve its scalability, this paper presents a low-rank-cum-Markov approximation (LMA) of the GP model that is novel in leveraging the dual computational advantages stemming from complementing a low-rank approximate representation of the full-rank GP based on a support set of inputs with a Markov approximation of the resulting residual process; the latter approximation is guaranteed to be closest in the Kullback-Leibler distance criterion subject to some constraint and is considerably more refined than that of existing sparse GP models utilizing low-rank representations due to its more relaxed conditional independence assumption (especially with larger data). As a result, our LMA method can trade off between the size of the support set and the order of the Markov property to (a) incur lower computational cost than such sparse GP models while achieving predictive performance comparable to them and (b) accurately represent features/patterns of any scale. Interestingly, varying the Markov order produces a spectrum of LMAs with PIC approximation and full-rank GP at the two extremes. An advantage of our LMA method is that it is amenable to parallelization on multiple machines/cores, thereby gaining greater scalability. Empirical evaluation on three real-world datasets in clusters of up to 32 computing nodes shows that our centralized and parallel LMA methods are significantly more time-efficient and scalable than state-of-the-art sparse and full-rank GP regression methods while achieving comparable predictive performances.

stat.ML↗

arXiv2014

Randomized Minmax Regret for Combinatorial Optimization Under Uncertainty

The minmax regret problem for combinatorial optimization under uncertainty can be viewed as a zero-sum game played between an optimizing player and an adversary, where the optimizing player selects a solution and the adversary selects costs with the intention of maximizing the regret of the player. The existing minmax regret model considers only deterministic solutions/strategies, and minmax regret versions of most polynomial solvable problems are NP-hard. In this paper, we consider a randomized model where the optimizing player selects a probability distribution (corresponding to a mixed strategy) over solutions and the adversary selects costs with knowledge of the player's distribution, but not its realization. We show that under this randomized model, the minmax regret version of any polynomial solvable combinatorial problem becomes polynomial solvable. This holds true for both the interval and discrete scenario representations of uncertainty. Using the randomized model, we show new proofs of existing approximation algorithms for the deterministic model based on primal-dual approaches. Finally, we prove that minmax regret problems are NP-hard under general convex uncertainty.

cs.DM↗

arXiv2014

Decentralized Data Fusion and Active Sensing with Mobile Sensors for Modeling and Predicting Spatiotemporal Traffic Phenomena

The problem of modeling and predicting spatiotemporal traffic phenomena over an urban road network is important to many traffic applications such as detecting and forecasting congestion hotspots. This paper presents a decentralized data fusion and active sensing (D2FAS) algorithm for mobile sensors to actively explore the road network to gather and assimilate the most informative data for predicting the traffic phenomenon. We analyze the time and communication complexity of D2FAS and demonstrate that it can scale well with a large number of observations and sensors. We provide a theoretical guarantee on its predictive performance to be equivalent to that of a sophisticated centralized sparse approximation for the Gaussian process (GP) model: The computation of such a sparse approximate GP model can thus be parallelized and distributed among the mobile sensors (in a Google-like MapReduce paradigm), thereby achieving efficient and scalable prediction. We also theoretically guarantee its active sensing performance that improves under various practical environmental conditions. Empirical evaluation on real-world urban road network data shows that our D2FAS algorithm is significantly more time-efficient and scalable than state-oftheart centralized algorithms while achieving comparable predictive performance.

cs.AI↗

arXiv2014

Parallel Gaussian Process Regression with Low-Rank Covariance Matrix Approximations

Gaussian processes (GP) are Bayesian non-parametric models that are widely used for probabilistic regression. Unfortunately, it cannot scale well with large data nor perform real-time predictions due to its cubic time cost in the data size. This paper presents two parallel GP regression methods that exploit low-rank covariance matrix approximations for distributing the computational load among parallel machines to achieve time efficiency and scalability. We theoretically guarantee the predictive performances of our proposed parallel GPs to be equivalent to that of some centralized approximate GP regression methods: The computation of their centralized counterparts can be distributed among parallel machines, hence achieving greater time efficiency and scalability. We analytically compare the properties of our parallel GPs such as time, space, and communication complexity. Empirical evaluation on two real-world datasets in a cluster of 20 computing nodes shows that our parallel GPs are significantly more time-efficient and scalable than their centralized counterparts and exact/full GP while achieving predictive performances comparable to full GP.

cs.LG↗

arXiv2014

Advances on Matroid Secretary Problems: Free Order Model and Laminar Case

The most well-known conjecture in the context of matroid secretary problems claims the existence of a constant-factor approximation applicable to any matroid. Whereas this conjecture remains open, modified forms of it were shown to be true, when assuming that the assignment of weights to the secretaries is not adversarial but uniformly random (Soto [SODA 2011], Oveis Gharan and Vondrák [ESA 2011]). However, so far, there was no variant of the matroid secretary problem with adversarial weight assignment for which a constant-factor approximation was found. We address this point by presenting a 9-approximation for the \emph{free order model}, a model suggested shortly after the introduction of the matroid secretary problem, and for which no constant-factor approximation was known so far. The free order model is a relaxed version of the original matroid secretary problem, with the only difference that one can choose the order in which secretaries are interviewed. Furthermore, we consider the classical matroid secretary problem for the special case of laminar matroids. Only recently, a constant-factor approximation has been found for this case, using a clever but rather involved method and analysis (Im and Wang, [SODA 2011]) that leads to a 16000/3-approximation. This is arguably the most involved special case of the matroid secretary problem for which a constant-factor approximation is known. We present a considerably simpler and stronger $3\sqrt{3}e\approx 14.12$-approximation, based on reducing the problem to a matroid secretary problem on a partition matroid.

cs.DS↗

arXiv2014

A Decomposition Algorithm for Nested Resource Allocation Problems

We propose an exact polynomial algorithm for a resource allocation problem with convex costs and constraints on partial sums of resource consumptions, in the presence of either continuous or integer variables. No assumption of strict convexity or differentiability is needed. The method solves a hierarchy of resource allocation subproblems, whose solutions are used to convert constraints on sums of resources into bounds for separate variables at higher levels. The resulting time complexity for the integer problem is $O(n \log m \log (B/n))$, and the complexity of obtaining an $ε$-approximate solution for the continuous case is $O(n \log m \log (B/ε))$, $n$ being the number of variables, $m$ the number of ascending constraints (such that $m < n$), $ε$ a desired precision, and $B$ the total resource. This algorithm attains the best-known complexity when $m = n$, and improves it when $\log m = o(\log n)$. Extensive experimental analyses are conducted with four recent algorithms on various continuous problems issued from theory and practice. The proposed method achieves a higher performance than previous algorithms, addressing all problems with up to one million variables in less than one minute on a modern computer.

cs.DS↗

arXiv2013

Average-Case Performance of Rollout Algorithms for Knapsack Problems

Rollout algorithms have demonstrated excellent performance on a variety of dynamic and discrete optimization problems. Interpreted as an approximate dynamic programming algorithm, a rollout algorithm estimates the value-to-go at each decision stage by simulating future events while following a greedy policy, referred to as the base policy. While in many cases rollout algorithms are guaranteed to perform as well as their base policies, there have been few theoretical results showing additional improvement in performance. In this paper we perform a probabilistic analysis of the subset sum problem and knapsack problem, giving theoretical evidence that rollout algorithms perform strictly better than their base policies. Using a stochastic model from the existing literature, we analyze two rollout methods that we refer to as the consecutive rollout and exhaustive rollout, both of which employ a simple greedy base policy. For the subset sum problem, we prove that after only a single iteration of the rollout algorithm, both methods yield at least a 30% reduction in the expected gap between the solution value and capacity, relative to the base policy. Analogous results are shown for the knapsack problem.

cs.DS↗

arXiv2013

Digital breadcrumbs: Detecting urban mobility patterns and transport mode choices from cellphone networks

Many modern and growing cities are facing declines in public transport usage, with few efficient methods to explain why. In this article, we show that urban mobility patterns and transport mode choices can be derived from cellphone call detail records coupled with public transport data recorded from smart cards. Specifically, we present new data mining approaches to determine the spatial and temporal variability of public and private transportation usage and transport mode preferences across Singapore. Our results, which were validated by Singapore's quadriennial Household Interview Travel Survey (HITS), revealed that there are 3.5 (HITS: 3.5 million) million and 4.3 (HITS: 4.4 million) million inter-district passengers by public and private transport, respectively. Along with classifying which transportation connections are weak or underserved, the analysis shows that the mode share of public transport use increases from 38 percent in the morning to 44 percent around mid-day and 52 percent in the evening.

cs.SI↗

arXiv2013

Greedy Online Bipartite Matching on Random Graphs

We study the average performance of online greedy matching algorithms on $G(n,n,p)$, the random bipartite graph with $n$ vertices on each side and edges occurring independently with probability $p=p(n)$. In the online model, vertices on one side of the graph are given up front while vertices on the other side arrive sequentially; when a vertex arrives its edges are revealed and it must be immediately matched or dropped. We begin by analyzing the \textsc{oblivious} algorithm, which tries to match each arriving vertex to a random neighbor, even if the neighbor has already been matched. The algorithm is shown to have a performance ratio of at least $1-1/e$ for all monotonic functions $p(n)$, where the performance ratio is defined asymptotically as the ratio of the expected matching size given by the algorithm to the expected maximum matching size. Next we show that the conventional \textsc{greedy} algorithm, which assigns each vertex to a random unmatched neighbor, has a performance ratio of at least 0.837 for all monotonic functions $p(n)$. Under the $G(n,n,p)$ model, the performance of \textsc{greedy} is equivalent to the performance of the well known \textsc{ranking} algorithm, so our results show that \textsc{ranking} has a performance ratio of at least 0.837. We finally consider vertex-weighted bipartite matching. Our proofs are based on simple differential equations that describe the evolution of the matching process.

cs.DS↗

arXiv2013

Parallel Gaussian Process Regression with Low-Rank Covariance Matrix Approximations

stat.ML↗

arXiv2013

Log-Quadratic Bounds for the Gaussian Q-function

We present bounds of quadratic form for the logarithm of the Gaussian Q-function. We also show an analytical method for deriving log-quadratic approximations of the Q-function and give an approximation with absolute error less than $10^{-3}$.

math.PR↗