Search

Today’s advanced Reinforcement Learning algorithms produce black-box policies, that are often difficult to interpret and trust for a person. We introduce a policy distilling algorithm, building on the CN2 rule mining algorithm, that distills the policy into a rule-based decision system. At the core of our approach is the fact that an RL process does not just learn a policy, a mapping from states to actions, but also produces extra meta-information, such as action values indicating the quality of alternative actions. This meta-information can indicate whether more than one action is near-optimal for a certain state. We extend CN2 to make it able to leverage knowledge about equally-good actions to distill the policy into fewer rules, increasing its interpretability by a person. Then, to ensure that the rules explain a valid, non-degenerate policy, we introduce a refinement algorithm that fine-tunes the rules to obtain good performance when executed in the environment. We demonstrate the applicability of our algorithm on the Mario AI benchmark, a complex task that requires modern reinforcement learning algorithms including neural networks. The explanations we produce capture the learned policy in only a few rules, that allow a person to understand what the black-box agent learned. Source code: https://gitlab.ai.vub.ac.be/yocoppen/svcn2., info:eu-repo/semantics/published

We have defined an interdisciplinary program for training a new generation of researchers who will be ready to leverage the use of artificial intelligence (ai)-based models and techniques even by non-expert users. The final goal is to make ai self-explaining and thus contribute to translating knowledge into products and services for economic and social benefit, with the support of explainable ai systems. Moreover, our focus is on the automatic generation of interactive explanations in natural language, the preferred modality among humans, with visualization as a complementary modality.keywordsexplainable artificial intelligencetrustworthinessmulti-modal explanationsargumentative conversational agentshuman-centered modelinghuman-machine persuasive interaction.

The problem of building a team to perform a complex task is often more than an optimal assignment of subtasks to agents based on individual performances. Subtasks may have subtle dependencies and relations that affect the overall performance of the formed team. This paper investigates the dependencies between subtasks and introduces some desired qualities of teams, such as preserving privacy or fairness. It proposes algorithms to analyze and build teams by taking into account the dependencies of assigned subtasks and agent performances. The performance of the algorithms are evaluated experimentally based on a multiagent system that is developed to answer complex queries. We show that by improving an initial team iteratively, the algorithm obtains teams with higher performance.

This is the era of big-data: high-volume, high-velocity and high-variety information assets are being collected, demanding cost-effective information processing. Analytic techniques primarily based on statistical methods are showing astonishing results, but exhibit also limited reasoning capabilities. On the other end of the spectrum the era of big- reasoning is emerging with next-generation cognitive and autonomous end-To-end solvers. A problem description in terms of text and diagrams is given: problem solvers should automatically understand the problem, identify its components, devise a model, identify a solving technique and find a solution with no human intervention. We propose a challenge: To design and implement an end-To-end solver for mathematical puzzles able to compete with primary school students. Mathematical puzzles require mathematics to solve them, but also logic, intuition and imagination are essential ingredients, thus calling for an unprecedented integration of many different AI techniques.

This paper considers multi-stage optimization problems under uncertainty that involve distinct offline and online phases. In particular it addresses the issue of integrating these phases to show how the two are often interrelated in real-world applications. Our methods are applicable under two (fairly general) conditions: 1) the uncertainty is exogenous; 2) it is possible to define a greedy heuristic for the online phase that can be modeled as a parametric convex optimization problem. We start with a baseline composed by a two-stage offline approach paired with the online greedy heuristic. We then propose multiple methods to tighten the offline/online integration, leading to significant quality improvements, at the cost of an increased computation effort either in the offline or the online phase. Overall, our methods provide multiple options to balance the solution quality/time trade-off, suiting a variety of practical application scenarios. To test our methods, we ground our approaches on two real cases studies with both offline and online decisions: an energy management problem with uncertain renewable generation and demand, and a vehicle routing problem with uncertain travel times. The application domains feature respectively continuous and discrete decisions. An extensive analysis of the experimental results shows that indeed offline/online integration may lead to substantial benefits.

Regularization-based approaches for injecting constraints in Machine Learning (ML) were introducedto improve a predictive model via expert knowledge. Given the recent interest in ethical and trustworthy AI, however, several works are resorting to these approaches for the (conceptually different) task of enforcing desired properties over a ML model, e.g. fairness. When regularized methods are used to enforce constraints, a typical approach consists in adjusting the regularizer multipliersuntil a suitable compromise between accuracy and constraint satisfaction is reached(e.g. a discrimination index becomes sufficiently low). This approach enables the use of traditional training algorithms, at the cost of having to search over the space of possible multipliers.Though the method is known to work well in many practical cases, the process has been subject to little general analysis. With this note, we make a preliminary step in this direction, providing a more systematic overview of the strengths and (in particular) potential weaknesses of this class of approaches. &nbsp

In data mining, finding interesting patterns is a challenging task. Constraint-based mining is a well-known approach to this, and one for which constraint programming has been shown to be a well-suited and generic framework. Constraint dominance programming (CDP) has been proposed as an extension that can capture an even wider class of constraint-based mining problems, by allowing us to compare relations between patterns. In this paper we improve CDP with the ability to specify an incomparability condition. This allows us to overcome two major shortcomings of CDP: finding dominated solutions that must then be filtered out after search, and unnecessarily adding dominance blocking constraints between incomparable solutions. We demonstrate the efficacy of our approach by extending the problem specification language ESSENCE and implementing it in a solver-independent manner on top of the constraint modelling tool CONJURE. Our experiments on pattern mining tasks with both a CP solver and a SAT solver show that using the incomparability condition during search significantly improves the efficiency of dominance programming and reduces (and often eliminates entirely) the need for post-processing to filter dominated solutions. Publisher PDF

One of the biggest challenges in the design of real-world decision support systems is coming up with a good combinatorial optimization model. Often enough, accurate predictive models (e.g. simulators) can be devised, but they are too complex or too slow to be employed in combinatorial optimization. In this paper, we propose a methodology called Empirical Model Learning (EML) that relies on Machine Learning for obtaining components of a prescriptive model, using data either extracted from a predictive model or harvested from a real system. In a way, EML can be considered as a technique to merge predictive and prescriptive analytics. All models introduce some form of approximation. Citing G.E.P. Box [1] “Essentially, all models are wrong, but some of them are useful”. In EML, models are useful if they provide adequate accuracy, and if they can be effectively exploited by solvers for finding high-quality solutions. We show how to ground EML on a case study of thermal-aware workload dispatching. We use two learning methods, namely Artificial Neural Networks and Decision Trees and we show how to encapsulate the learned model in a number of optimization techniques, namely Local Search, Constraint Programming, Mixed Integer Non-Linear Programming and SAT Modulo Theories. We demonstrate the effectiveness of the EML approach by comparing our results with those obtained using expert-designed models.

In the past few years, the area of Machine Learning (ML) has witnessed tremendous advancements, becoming a pervasive technology in a wide range of applications. One area that can significantly benefit from the use of ML is Combinatorial Optimization. The three pillars of constraint satisfaction and optimization problem solving, i.e., modeling, search, and optimization, can exploit ML techniques to boost their accuracy, efficiency and effectiveness. In this survey we focus on the modeling component, whose effectiveness is crucial for solving the problem. The modeling activity has been traditionally shaped by optimization and domain experts, interacting to provide realistic results. Machine Learning techniques can tremendously ease the process, and exploit the available data to either create models or refine expert-designed ones. In this survey we cover approaches that have been recently proposed to enhance the modeling process by learning either single constraints, objective functions, or the whole model. We highlight common themes to multiple approaches and draw connections with related fields of research., Comment: Originally submitted to IJCAI2018

In this work we present two general techniques to deal with multi-stage optimization problems under uncertainty, featuring off-line and on-line decisions. The methods are applicable when: 1) the uncertainty is exogenous; 2) there exists a heuristic for the on-line phase that can be modeled as a parametric convex optimization problem. The first technique replaces the on-line heuristics with an anticipatory solver, obtained through a systematic procedure. The second technique consists in making the off-line solver aware of the on-line heuristic, and capable of controlling its parameters so as to steer its behavior. We instantiate our approaches on two case studies: an energy management system with uncertain renewable generation and load demand, and a vehicle routing problem with uncertain travel times. We show how both techniques achieve high solution quality w.r.t. an oracle operating under perfect information, by obtaining different trade-offs in terms of computation time.

Kidney exchange programs enable willing, but incompatible, donor-patient pairs to swap donors, thus allowing persons suffering from organ failure to access transplantation. Choosing which pairs to match requires solving a stochastic online optimization problem where patients and donors arrive over time. Despite this, most of the related scientific literature has focused on deterministic offline models. In this paper, we present a simple approach to employ a model for the offline Kidney Exchange Problem (KEP) as the basis of an on-line anticipatory algorithm. Our approach grounds on existing techniques for the on-line KEP, but it generalizes them and provides a more accurate estimate of the expected impact of current decisions. In an experimentation based on a state-of-the-art donor pool generation method, the approach provides improvements in terms of quality and is able to deal with realistic instance size in reasonable time.

none 4 no Deep Neural Networks (DNNs) have been shaking the AI scene, for their ability to excel at Machine Learning tasks without relying on complex, hand-crafted, features. Here, we probe whether a DNN can learn how to construct solutions of a CSP, without any explicit symbolic information about the problem constraints. We train a DNN to extend a feasible solution by making a single, globally consistent, variable assignment. The training is done over intermediate steps of the construction of feasible solutions. From a scientific standpoint, we are interested in whether a DNN can learn the structure of a combinatorial problem, even when trained on (arbitrarily chosen) construction sequences of feasible solutions. In practice, the network could also be used to guide a search process, e.g. to take into account (soft) constraints that are implicit in past solutions or hard to capture in a traditional declarative model. This research line is still at an early stage, and a number of complex issues remain open. Nevertheless, we already have intriguing results on the classical Partial Latin Square and N-Queen completion problems. mixed Galassi Andrea, Lombardi Michele, Mello Paola, Milano Michela Galassi Andrea, Lombardi Michele, Mello Paola, Milano Michela

Policy making is an extremely complex process occurring in changing environments and affecting the three pillars of sustainable development: society, economy and the environment. Each political decision in fact implies some form of social reactions, it affects economic and financial aspects and has substantial environmental impacts. Improving decision making in this context could have a huge beneficial impact on all these aspects. There are a number of Artificial Intelligence techniques that could play an important role in improving the policy making process such as decision support and optimization techniques, game theory, data and opinion mining and agent-based simulation. We outline here some potential use of AI technology as it emerged by the European Union (EU) EU FP7 project ePolicy: Engineering the Policy Making Life-Cycle, and we identify some potential research challenges.

Recently, a number of noteworthy results have been achieved in various fields of artificial intelligence, and many aspects of the problem solving process have received significant attention by the scientific community. In this context, the extraction of comprehensive knowledge suitable for problem solving and reasoning, from textual and pictorial problem descriptions, has been less investigated, but recognized as essential for autonomous thinking in Artificial Intelligence. In this work we present a challenge where methods and tools for deep understanding are strongly needed for enabling problem solving: we propose to solve mathematical puzzles by means of computers, starting from text and diagrams describing them, without any human intervention. We are aware that the proposed challenge is hard and of difficult solution nowadays (and in the foreseeable future), but even studying and solving only single parts of the proposed challenge would represent an important step forward for artificial intelligence.

I started my term as Editor in Chief of the Constraint Journal on the 1st of January 2015. The purpose of this short report is to provide some news about the journal to the Editorial and Advisory board members and to the constraint programming community.

Scheduling and dispatching are critical enabling technologies in supercomputing and grid computing. In these contexts, scalability is an issue: we have to allocate and schedule upi?źto tens of thousands of tasks on tens of thousands of resources. This problem scale is out of reach for complete and centralized scheduling approaches. We propose a distributed allocation and scheduling paradigm called DARDIS that is lightweight, scalable and fully customizable in many domains. In DARDIS each task offloads to the available resources the computation of a probability index associated with each possible start time for the given task on the specific resource. The task then selects the proper resource and start time on the basis of the above probability. The scheduler can be customized with different policies to fit several objective functions like load balancing or makespan. We evaluate our approach in the domain of grids and supercomputers. We compare DARDIS with the most widely used algorithms used in these specific domains to show that this approach can reach better solutions in several cases.

Our society is organized around a number of (interdependent) global systems. Logistic and supply chains, health services, energy networks, financial markets, computer networks, and cities are just a few examples of such global, complex systems. These global systems are socio-technical and involve interactions between complex infrastructures, man-made processes, natural phenomena, multiple stakeholders, and human behavior. For the first time in the history of manking, we have access to data sets of unprecedented scale and accuracy about these infrastructures, processes, natural phenomena, and human behaviors. In addition, progress in high-performancing computing, data mining, machine learning, and decision support opens the possibility of looking at these problems more holistically, capturing many of these aspects simultaneously. This paper addresses emergent architectures enabling controlling, predicting and reaoning on these systems.

This paper presents a Traffic Lights control system, inspired by Swarm intelligence methodologies, in which every intersection controller makes independent decisions to pursue common goals and is able to improve global traffic performance. The solution is low cost and widely applicable to different urban scenarios. This work is developed within the COLOMBO European project. Control methods are divided into macroscopic and microscopic control level methods. The former react to macroscopic key figures such as mean congestion length and mean traffic density and acts on the signal program choice or the development of the frame signal program. The latter include changes at short notice based on changes in the traffic flow: they include methods for signal program adaptation and development. The developed system has been widely tested on synthetic benchmarks with promising results.

In this paper, we propose a challenging research direction for Constraint Programming and optimization techniques in general. We address problems where decisions to be taken affect and are affected by complex systems, which exhibit phenomena emerging from a collection of interacting objects, capable to self organize and to adapt their behaviour according to their history and feedback. Such systems are unfortunately impervious to modeling efforts via state-of-the-art combinatorial optimization techniques. We provide some hints on approaches to connect and integrate decision making and optimization technology with complex systems via machine learning, game theory and mechanism design. In the first case, the aim is to extract modeling components to express the relation between global decisions and observables emerging from the real system, or from an accurate predictive model (e.g. a simulator). In the second case, the idea is to exploit game theory, mechanism design and distributed decision making to drive the process toward realistic equilibrium points avoiding globally optimal, but unrealistic, configurations. We conclude by observing how dealing with the complexity of the considered problems will require to greatly extend the capabilities of state of the art solvers: in this context, we identify some key issues and highlight future research directions.

An interesting exercise for a research community is to foresee its future, identify key research challenges and open problems and be forward-looking in both technology and application fields. The Constraint Programming community, in more than 20 years, has performed this exercise several times. One example is the recent book on Trends in Constraint Programming edited in 2007 by Benhamou, Jussien and O’Sullivan [1] and the recent challenge paper at AAAI 2012 by O’Sullivan [17] entitled Opportunities and challenges for Constraint Programming. Other examples are the earlier papers by Van Hentenryck and Saraswat [8] on Strategic Directions in Constraint Programming in 1996 and the ACM 50th Anniversary Issue of ACM Computing Surveys of the same year on Strategic Directions in Computing Research containing a number of forward looking papers [2–7, 9–16, 18–21]. Some of these papers have correctly foreseen the future, some challenges have received less then expected attention, but overall these papers have always provided a useful guide for researchers and newcomers. Here we are offering another snapshot of the constraint programming community and what its members consider as promising directions. We have intentionally invited contribution without structuring them into a schema. Rather we have preferred a bottom-up call for proposals providing supported-by-evidence viewpoints on future directions of CP.

This contribution is aimed at providing an overview of the main Italian research areas and activities. In this chapter we first analyze the collaboration structure of Italian research, that involves more than eight hundred scholars and researchers from both the University and the Industry and, from a network perspective, it appears to be scale-free. Then, we briefly illustrate the main subjects of investigation and applications. AI research in Italy traces back to the '70s with an increase in the last twenty years and spans the main research AI areas, from automated reasoning and ontologies, to machine learning, robotics and evolutionary computation.

International audience; The constraint NValue counts the number of different values assigned to a vector of variables. Propagating generalized arc consistency on this constraint is NP-hard. We show that computing even the lower bound on the number of values is NP-hard. We therefore study different approximation heuristics for this problem. We introduce three new methods for computing a lower bound on the number of values. The first two are based on the maximum independent set problem and are incomparable to a previous approach based on intervals. The last method is a linear relaxation of the problem. This gives a tighter lower bound than all other methods, but at a greater asymptotic cost.

Anomaly detection in supercomputers is a very difficult problem due to the big scale of the systems and the high number of components. The current state of the art for automated anomaly detection employs Machine Learning methods or statistical regression models in a supervised fashion, meaning that the detection tool is trained to distinguish among a fixed set of behaviour classes (healthy and unhealthy states). We propose a novel approach for anomaly detection in High Performance Computing systems based on a Machine (Deep) Learning technique, namely a type of neural network called autoencoder. The key idea is to train a set of autoencoders to learn the normal (healthy) behaviour of the supercomputer nodes and, after training, use them to identify abnormal conditions. This is different from previous approaches which where based on learning the abnormal condition, for which there are much smaller datasets (since it is very hard to identify them to begin with). We test our approach on a real supercomputer equipped with a fine-grained, scalable monitoring infrastructure that can provide large amount of data to characterize the system behaviour. The results are extremely promising: after the training phase to learn the normal system behaviour, our method is capable of detecting anomalies that have never been seen before with a very good accuracy (values ranging between 88% and 96%)., 9 pages, 3 figures