Asymmetries in Adversarial Settings

Even without formal guarantees of their effectiveness, adversarial attacks against Machine Learning models frequently fool new defenses. We identify six key asymmetries that contribute to this phenomenon and formulate four guidelines to build future-proof defenses by preventing such asymmetries. We also prove that attacking a classifier is NP-complete, while defending from such attacks is Sigma_2^P-complete. We then introduce Counter-Attack (CA), an asymmetry-free metadefense that determines whether a model is robust on a given input by estimating its distance from the decision boundary. Under specific assumptions CA can provide theoretical detection guarantees. Additionally, we prove that while CA is NP-complete, fooling CA is Sigma_2^P-complete. Even when using heuristic relaxations, we show that our method can reliably identify non-robust points. As part of our experimental evaluation, we introduce UG100, a new dataset obtained by applying a provably optimal attack to six limited-scale networks (three for MNIST and three for CIFAR10), each trained in three different manners.

Design and management of a supply chain: a joint location-inventory problem under uncertainty (Part A)

In this paper, a joint location-inventory model is proposed that simultaneously optimises strategic supply chain design decisions such as facility location and customer allocation to facilities, and tactical-operational inventory management and production scheduling decisions. All this is analysed in a context of demand uncertainty and supply uncertainty. While demand uncertainty stems from potential fluctuations in customer demands over time, supply-side uncertainty is associated with the risk of “disruption” to which facilities may be subject. The latter is caused by external factors such as natural disasters, strikes, changes of ownership and information technology security incidents. The proposed model is formulated as a non-linear mixed integer programming problem to minimise the expected total cost, which includes four basic cost items: the fixed cost of locating facilities at candidate sites, the cost of transport from facilities to customers, the cost of working inventory, and the cost of safety stock. Next, since the optimisation problem is very complex and the number of evaluable instances is very low, a "matheuristic" solution is presented. This approach has a twofold objective: on the one hand, it considers a larger number of facilities and customers within the network in order to reproduce a supply chain configuration that more closely reflects a real-world context; on the other hand, it serves to generate a starting solution and perform a series of iterations to try to improve it. Thanks to this algorithm, it was possible to obtain a solution characterised by a lower total system cost than that observed for the initial solution. The study concludes with some reflections and the description of possible future insights.

Fleet Management System for Search and Rescue Operations

Over one million people lost their lives in the last twenty years from natural disasters like wildfires, earthquakes and man-made disasters. In such scenarios the usage of a fleet of robots aims at the parallelization of the workload and thus increasing speed and capabilities to complete time sensitive missions. This work focuses on the development of a dynamic fleet management system, which consists in the management of multiple agents cooperating in order to accomplish tasks. We presented a Mixed Integer Programming problem for the management and planning of mission’s tasks. The problem was solved using both an exact and a heuristic approach. The latter is based on the idea of solving iteratively smaller instances of the complete problem. Alongside, a fast and efficient algorithm for estimation of travel times between tasks is proposed. Experimental results demonstrate that the proposed heuristic approach is able to generate quality solutions, within specific time limits, compared to the exact one.

Using sentinels to detect intersections of convex and nonconvex polygons

We describe finite sets of points, called sentinels, which allow us to decide if isometric copies of polygons, convex or not, intersect. As an example of the applicability of the concept of sentinel, we explain how they can be used to formulate an algorithm based on the optimization of differentiable models to pack polygons in convex sets. Mathematical subject classification: 90C53, 65K05.

A STUDY OF PENALTY FORMULATIONS USED IN THE NUMERICAL APPROXIMATION OF A RADIALLY SYMMETRIC ELASTICITY PROBLEM

We consider a class of two-dimensional problems in classical linear elasticity for which material overlapping occurs in the absence of singularities. Of course, material overlapping is not physically realistic, and one possible way to prevent it uses a constrained minimization theory. In this theory, a minimization problem consists of minimizing the total potential energy of a linear elastic body subject to the constraint that the deformation field must be locally invertible. Here, we use an interior and an exterior penalty formulation of the minimization problem together with both a standard finite element method and classical nonlinear programming techniques to compute the minimizers. We compare both formulations by solving a plane problem numerically in the context of the constrained minimization theory. The problem has a closed-form solution, which is used to validate the numerical results. This solution is regular everywhere, including the boundary. In particular, we show numerical results which indicate that, for a fixed finite element mesh, the sequences of numerical solutions obtained with both the interior and the exterior penalty formulations converge to the same limit function as the penalization is enforced. This limit function yields an approximate deformation field to the plane problem that is locally invertible at all points in the domain. As the mesh is refined, this field converges to the exact solution of the plane problem.

A systemic approach for optimal cooling tower operation

The thermal performance of a cooling tower and its cooling water system is critical for industrial plants, and small deviations from the design conditions may cause severe instability in the operation and economics of the process. External disturbances such as variation in the thermal demand of the process or oscillations in atmospheric conditions may be suppressed in multiple ways. Nevertheless, such alternatives are hardly ever implemented in the industrial operation due to the poor coordination between the utility and process sectors. The complexity of the operation increases because of the strong interaction among the process variables. In the present work, an integrated model for the minimization of the operating costs of a cooling water system is developed. The system is composed of a cooling tower as well as a network of heat exchangers. After the model is verified, several cases are studied with the objective of determining the optimal operation. It is observed that the most important operational resources to mitigate disturbances in the thermal demand of the process are, in this order: the increase in recycle water flow rate, the increase in air flow rate and finally the forced removal of a portion of the water flow rate that enters the cooling tower with the corresponding make-up flow rate. (C) 2009 Elsevier Ltd. All rights reserved.

Investment optimization in distribution network based on fuzzy outage parameters

This paper presents a methodology that aims to increase the probability of delivering power to any load point of the electrical distribution system by identifying new investments in distribution components. The methodology is based on statistical failure and repair data of the distribution power system components and it uses fuzzy-probabilistic modelling for system component outage parameters. Fuzzy membership functions of system component outage parameters are obtained by statistical records. A mixed integer non-linear optimization technique is developed to identify adequate investments in distribution networks components that allow increasing the availability level for any customer in the distribution system at minimum cost for the system operator. To illustrate the application of the proposed methodology, the paper includes a case study that considers a real distribution network.

Solving MPCC problem with the hyperbolic penalty function

The main goal of this work is to solve mathematical program with complementarity constraints (MPCC) using nonlinear programming techniques (NLP). An hyperbolic penalty function is used to solve MPCC problems by including the complementarity constraints in the penalty term. This penalty function [1] is twice continuously differentiable and combines features of both exterior and interior penalty methods. A set of AMPL problems from MacMPEC [2] are tested and a comparative study is performed.

Inexact restoration approaches to solve mathematical program with complementarity constraints

Mathematical Program with Complementarity Constraints (MPCC) finds applica- tion in many fields. As the complementarity constraints fail the standard Linear In- dependence Constraint Qualification (LICQ) or the Mangasarian-Fromovitz constraint qualification (MFCQ), at any feasible point, the nonlinear programming theory may not be directly applied to MPCC. However, the MPCC can be reformulated as NLP problem and solved by nonlinear programming techniques. One of them, the Inexact Restoration (IR) approach, performs two independent phases in each iteration - the feasibility and the optimality phases. This work presents two versions of an IR algorithm to solve MPCC. In the feasibility phase two strategies were implemented, depending on the constraints features. One gives more importance to the complementarity constraints, while the other considers the priority of equality and inequality constraints neglecting the complementarity ones. The optimality phase uses the same approach for both algorithm versions. The algorithms were implemented in MATLAB and the test problems are from MACMPEC collection.

Funções penalidade para o tratamento das variáveis discretas do problema de fluxo de potência ótimo reativo

Conselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq)

Leilões combinados

Mestrado em Controlo de Gestão e dos Negócios

Spinning reserve and emission unit commitment through stochastic optimization

This paper proposes a stochastic mixed-integer linear approach to deal with a short-term unit commitment problem with uncertainty on a deregulated electricity market that includes day-ahead bidding and bilateral contracts. The proposed approach considers the typically operation constraints on the thermal units and a spinning reserve. The uncertainty is due to the electricity prices, which are modeled by a scenario set, allowing an acceptable computation. Moreover, emission allowances are considered in a manner to allow for the consideration of environmental constraints. A case study to illustrate the usefulness of the proposed approach is presented and an assessment of the cost for the spinning reserve is obtained by a comparison between the situation with and without spinning reserve.

Classification of some penalty methods

Optimization problems arise in science, engineering, economy, etc. and we need to find the best solutions for each reality. The methods used to solve these problems depend on several factors, including the amount and type of accessible information, the available algorithms for solving them, and, obviously, the intrinsic characteristics of the problem. There are many kinds of optimization problems and, consequently, many kinds of methods to solve them. When the involved functions are nonlinear and their derivatives are not known or are very difficult to calculate, these methods are more rare. These kinds of functions are frequently called black box functions. To solve such problems without constraints (unconstrained optimization), we can use direct search methods. These methods do not require any derivatives or approximations of them. But when the problem has constraints (nonlinear programming problems) and, additionally, the constraint functions are black box functions, it is much more difficult to find the most appropriate method. Penalty methods can then be used. They transform the original problem into a sequence of other problems, derived from the initial, all without constraints. Then this sequence of problems (without constraints) can be solved using the methods available for unconstrained optimization. In this chapter, we present a classification of some of the existing penalty methods and describe some of their assumptions and limitations. These methods allow the solving of optimization problems with continuous, discrete, and mixing constraints, without requiring continuity, differentiability, or convexity. Thus, penalty methods can be used as the first step in the resolution of constrained problems, by means of methods that typically are used by unconstrained problems. We also discuss a new class of penalty methods for nonlinear optimization, which adjust the penalty parameter dynamically.

Filters method in direct search optimization, new measures to admissibility

Constrained nonlinear optimization problems are usually solved using penalty or barrier methods combined with unconstrained optimization methods. Another alternative used to solve constrained nonlinear optimization problems is the lters method. Filters method, introduced by Fletcher and Ley er in 2002, have been widely used in several areas of constrained nonlinear optimization. These methods treat optimization problem as bi-objective attempts to minimize the objective function and a continuous function that aggregates the constraint violation functions. Audet and Dennis have presented the rst lters method for derivative-free nonlinear programming, based on pattern search methods. Motivated by this work we have de- veloped a new direct search method, based on simplex methods, for general constrained optimization, that combines the features of the simplex method and lters method. This work presents a new variant of these methods which combines the lters method with other direct search methods and are proposed some alternatives to aggregate the constraint violation functions.

Direct-search penalty/barrier methods

In Nonlinear Optimization Penalty and Barrier Methods are normally used to solve Constrained Problems. There are several Penalty/Barrier Methods and they are used in several areas from Engineering to Economy, through Biology, Chemistry, Physics among others. In these areas it often appears Optimization Problems in which the involved functions (objective and constraints) are non-smooth and/or their derivatives are not know. In this work some Penalty/Barrier functions are tested and compared, using in the internal process, Derivative-free, namely Direct Search, methods. This work is a part of a bigger project involving the development of an Application Programming Interface, that implements several Optimization Methods, to be used in applications that need to solve constrained and/or unconstrained Nonlinear Optimization Problems. Besides the use of it in applied mathematics research it is also to be used in engineering software packages.