The Doctoral Consortium will take place on July 19 and 20.
Chairs and organizers: Josep Puyol-Gruart, Jordi Sabater-Mir.
Committees: Eva Armengol, Filippo Bistaffa, Christian Blum, Jesus Cerquides, Lluís Godo, Dave de Jonge, Ramon Lopez de Mantaras, Oguz Mulayim, Pablo Noriega, Enric Plaza, Josep Puyol-Gruart, Jordi Sabater-Mir, Marco Schorlemmer.
Every student will have 15 minutes to explain his/her PhD progress and plans and 10 minutes more to discuss with the committee's three members. Students should bear in mind that the level of the explanation has to be understood by an audience with knowledge in AI but not in the specific area of interest and should give particular importance to the impact that the PhD can potentially have. Presentations are to be done in English.
Possible presentation squeme:
The coffees and the lunch of the students and members of the committee will be free (students will have free lunch the day of his presentation).
After the presentations, there will be a consensus meeting of the committee members to prepare the final reports.
The schedule of the presentations is the following:
Creation of a modular Decision Support System to improve results in Assisted Reproduction
Much like a lot of fields in healthcare, Assisted Reproduction generates lots of data that up until now has gone largely unused. Even after approximately 40 years of the first successful In Vitro Fertilization treatment and all the gigantic strides made to advance Assisted Reproduction Techniques until today, the chances of achieving pregnancy after In Vitro Fertilization remain around 30%. With the aim to improve that ratio, this research will focus on applying Artificial Intelligence and Machine Learning techniques to unfold the power of the high quantity of data generated by Assisted Reproduction Techniques. Specifically designing a modular Decision Support System aiming at improving efficacy after Controlled Ovarian Stimulation, and assessing pregnancy probabilities and multiple pregnancy risk after embryo transfer.
Moral values machine learning in multi-agent systems
The main objective of this thesis is to define formally what ethical behavior is in a multi-agent system, based on both ethical theory and game theory. Specifically we want to be able to formally define what it means to be aligned with a system of moral values in multi-agent environments such as Markov Games. When formalised, we want to provide an effective algorithm for a multi-agent system to learn to behave aligned with a system of moral values.
Developing Efficient Routing Algorithms for Sustainable City Logistics
Increasing environmental concerns and legal regulations have led to the development of sustainable technologies and systems in logistics, as in many fields. The adoption of multi-echelon distribution networks and environmentally-friendly vehicles in freight distribution have become major concepts for reducing the negative impact of urban transportation activities. In this line, this thesis addresses a two-echelon electric vehicle routing problem (2E-EVRP) as a practice of sustainable city logistics. In the first echelon of the distribution network, products are transported from central warehouses to satellites located in the surroundings of cities. This is achieved by means of large conventional trucks. Subsequently, relatively smaller-sized electric vehicles distribute these products from the satellites to demand points/customers in the cities as they are less noisy and have zero direct emission. The problem addressed in this study takes into account the limited driving range of electric vehicles that need to be recharged in charging stations when necessary. Furthermore, we considered realistic charging scenarios such as partial recharging and non-linear battery charging time and observed their effect on the solution quality. In addition, the proposed problem considers some of the essential real-life constraints, i.e., time window and simultaneous pickup and delivery. A mixed-integer linear programming formulation is developed, and small-sized instances are solved using CPLEX. Due to the complexity of the problem, we have developed solution algorithms based on variable neighborhood search and construct, merge, solve and adapt to solve large-sized problem instances.
Roger Xavier Lera Leri
Explainability for optimisation-based decision support
In the last years, there has been an increasing interest in developing Artificial Intelligence (AI) systems centered in humans. Thus, such AI systems must be trustworthy, meaning that they have to be ethical, lawful and robust. Within this new vision of AI, there is an strong consensus to require explainability in AI algorithms, i.e. the capacity to provide explanations of the decisions taken by such algorithms.
Hence, the goal of this PhD thesis is to develop decision support systems that not only recommend the optimal solutions of different real-world problems, but also to develop a general framework for explainable AI that provides explanations of the decisions taken by our approaches. We aim at formalising such problems as convex optimisation problems, enabling the use of commercial-off-the-shelf solvers to solve such problems and using real-world data instances to evaluate our approaches.
Machine Learning Platform for Assisted Reproductive Technologies
In-vitro fertilization (IVF) is a domain that faces large problems in efficiency and efficacy, for human patients and mammals. So far, even though there are critical advancements in the applications of Artificial Intelligence (AI) and Machine Learning (ML) in Healthcare, but also more specifically in assistive reproduction, the use of these kinds of techniques remains low. The reasons lie in different factors, from the unsatisfactory accuracy rates to the lack of mathematical evidence for the results of some applications, and the mistrust of correctness by Healthcare scientists and professionals. The objectives of this research are the design and development of software, based on AI and ML techniques, to further support the research in assisted reproductive technologies (ART) and answer some of the existing doubts on the current applications. The contribution of this project involves (i) the design of new machine learning algorithms, (ii) the facilitation of the acquisition, analysis and validation of new knowledge through the analysis of ART data, and (iii) the introduction of new, proven AI methods for the evaluation of IVF.
Engineering pro-social values in autonomous agents -- Collective and individual perspectives
In the field of multi-agent systems, a set of autonomous agents (possibly including humans) interact with one another and with their shared environment to compete for resources, collaborate in the completion a task, or negotiate an agreement. In order to ensure that these interactions are ethically compliant and mutually beneficial, the values that the participant agents and the multi-agent system as a whole should abide by have to: (a) be embedded into them, and (b) have a clear social orientation. The goal of this thesis, then, is to formulate and implement models to engineer pro-social values into systems composed of autonomous agents. To this end, two complementary approaches are being explored. First, the collective perspective takes the view of an external social planner who leverages prescriptive norms as an avenue to ensure that the system is compliant with respect to some values. Such values has been previously grounded, aka their semantics have been defined through state features that act as proxies for it. Second, the individual perspective looks to engineer pro-social values at the individual agent level. In order to ensure exclusively socially-oriented behaviour, agents are endowed with Theory of Mind capabilities, i.e. the ability to reason about the beliefs, desires, and intentions of others. We formulate and implement a cognitive agent model that is able to observe another agent's (the actor) action, put themselves in the shoes of the actor, and infer through abductive reasoning the knowledge the actor was relying upon when selecting his action. By means of this empathetic cognitive machinery, an individual agent is able to effectively engage in cooperation with other, equally pro-social agents.
Team Formation Methods for Dynamic Large-Scale Competence Based Problems
Team formation is the problem of building groups of individuals that need to collaborate in order to collectively undertake a given task. The problem has been studied in several scientific areas including computer sciences, economics, psychology, and social sciences. At the same time, team formation within companies and organizations is usually characterised as a hard and time consuming problem for managers and human resources. This thesis tackles the combinatorial problem of (a) forming teams, and (b) matching teams to tasks, especially as the number of individuals and the number of tasks increase. Specifically we intend to devise automated methods that ease the team formation process and yield sufficiently good results in large-scale scenarios. In order to do so, we consider competencies required by the tasks that match the competencies collectively acquired by the team; while we exploit features that has been evidenced to affect teams' performance such as preferences over tasks, preferences over potential partners (that drives the social cohesion within a team), and individuals’ personality and gender. Moreover, recognising the fact that humans tend to disbelieve or distrust an AI tool, we enhance our team formation tool to provide explanations.
Is This a Violation? Learning and Understanding Norm Violations in Online Communities
This work focuses on normative systems for online communities. We are interested in addressing the issue that arises when different community members interpret these norms in different ways, possibly leading to unexpected behavior in interactions, usually with norm violations that affect the individual and community experiences. The challenge is to learn the meaning of a norm violation from the limited interaction data and incorporate explainability to improve the community experience on understanding why an action was classified as a violation. For this, we use batch and incremental learning to train an ensemble of classifiers. Ensemble learning handles the imbalanced class distribution of the interaction stream, while the training approaches use different strategies to keep the ensemble models in accordance with the latest community view on the meaning of norm violation. Our proposed framework can handle featurized and text datasets. For the second case, we investigate the application of Natural Language Processing and transformers (specifically, at this initial stage, focusing on BERT).
A formal model of sense-making using image schemas and conceptual blending
Sense-making is a process seldom adressed in AI, while cognitive science approaches it as the process of an autonomous agent bringing its own original meaning upon its environment, and proposes it is fundamental for cognition. We therefore model the sense-making process as the conceptual blending of image schemas with a structural description of a stimulus. The case study we have used is diagrams and their geometric configurations. Image schemas comprise mental structures abstracting the invariances of repeated sensorimotor contingencies such as SUPPORT, VERTICALITY and BALANCE. They structure our perception and reasoning by transferring their structure to our percepts according to the principles of conceptual blending. In our work we model the conceptual blend of various image schemas with the geometry of a diagram, obtaining a blend that reflects the interpreted diagram. We formalize image schemas and geometric configurations with typed FOL theories, and, for the latter, use Qualitative Spatial Reasoning formalisms to describe the topology, shape, and relative position of its components. Conceptual blends are computed as category-theoretic colimits. The resulting blend has emergent structure, representing a meaningful diagram, e.g. The Hasse diagram (representing a poset) as a SCALE with levels, minimum and maximum elements etc. Our work on diagrams can provide guidelines for effective visualizations, and our general framework can be developed into a system that constructs possible conceptual meanings for various stimuli types.
Collective Intelligence in Intelligent Traffic Management Systems with Autonomous Vehicles
The major methodology I am using for my Ph.D. project is to treat each autonomous vehicle and smart roadside infrastructure as an intelligent agent or a robot so that the whole transport system becomes a multi-agent (multi-robot) system. My ultimate goal in the research is to find collective intelligence among the traffic agents/robots. I have introduced a spatial model of road network based on graph theory to represent the topological relationship of roads, which contains the connection of road lanes and intersections and the internal connections of an intersection. In addition, we have further implemented a generic simulator to test the game-theoretic properties of a transport system. Based on the generic simulation system, we can test a variety of properties of intelligent traffic management systems and autonomous vehicles from the macro to micro perspectives of transport networks. Currently, I am focusing on the properties of PoA in a transport system with purely AVs, which measures the gap in system costs between global optimization and self-decision making. I am trying to adopt the real-world traffic data from the selected Australian roads into our simulator so that I can estimate the PoA of a transport system by finding the system latency under equilibrium status and the optimal latency.
Social Simulation to Foster Collaborative Learning in Educational Settings
Given an educational context with real students in face-to-face and/or virtual classes, the goal of this PhD thesis is to generate an agent-based model of the class such that:
Considers and encourages the learning of each student individually. This implies a need to identify and differentiate students separately, each with a particular expertise level and pursuing a selfish goal: to learn as much as possible. In order to make such a distinction, an assessment of each student's knowledge is needed. This partially overlaps with the raison d'être of personalized learning.
Models group work and the osmosis of knowledge between members of a class. In opposition to the previous objective, it could be said that this one comes into direct confrontation with the backbone of adaptive or personalized learning. While personalized learning isolates the student to a certain extent by offering her a unique learning experience, we seek to model and find those peer interactions that will improve the cognitive representation that each individual student has of a particular learning topic. In other words: we aim to model how the understanding that a student has of a given topic is transmitted to her peers.
Comparison of the model's predictions with experimental data will allow us to refine, evaluate and validate our model. Should we arrive at a model whose outputs are acceptable with respect to the real data-generating process, this model could be used as an intelligent teaching supporting tool that informs optimal partitions/teaching policies in a class for a specific collaborative task.
Improving Optimization Algorithms with Machine Learning and With Graphical Tools
This thesis will deal principally with a recent line of research in optimization, which explores the potential of integrating machine learning (ML) techniques (in particular, deep learning) with optimization algorithms. In this project we are especially interested in the use of deep learning techniques for the discovery of new methods for the generation of valid solutions to hard combinatorial optimization problems. In other words, our goal is to make use of deep learning to learn how to generate good solutions for a (general) problem, as opposed to learning how to generate good solutions for a specific instance of a (particular) problem. Another topic that will be studied in this thesis is improving our understanding of the behaviour of optimization algorithms such as metaheuristics. In general, there is a lack of accessible tools to analyse, contrast and visualise the behaviour of metaheuristics when solving optimization problems. To help to answer these questions, a recent tool called search trajectory networks (STNs) was presented in the related literature. STNs is a data-driven, graph-based tool in order to analyse, visualise and directly contrast the behaviour of different types of metaheuristics. However, the usability of this tool is still very much limited. One of the goals of this thesis is to improve the initial STNs tool and to improve its usability.
Argumentation-based Multiagent Recommender System
The goal of this project is to develop a a multi-agent recommender system in which multiple agents collaborate with each other to finalise the recommendation for a user. During this process, they should generate a trace of the decision making process which can then be used to provide justifications for the recommendation to the user in an interactive manner.
An Algorithmic Framework for Making Use of Negative Learning in Ant Colony Optimization
Ant colony optimization is a metaheuristic that is mainly used for solving hard combinatorial optimization problems. The distinctive feature of ant colony optimization is a learning mechanism that is based on learning from positive examples. This is also the case in other learning-based metaheuristics such as evolutionary algorithms and particle swarm optimization. Examples from nature, however, indicate that negative learning—in addition to positive learning can beneficially be used for certain purposes. Several research papers have explored this topic over the last decades in the context of ant colony optimization, mostly with limited success. In this work, we study an alternative mechanism by making use of mathematical programming for the incorporation of negative learning in ant colony optimization. Moreover, we compare our proposal to some well-known existing negative learning approaches from the related literature. Currently we have implemented our approach to five types of subset selection problems: minimum dominating set, capacitated minimum dominating set, minimum positive influence dominating set, multi-dimensional knapsack, and maximum satisfiability. In all of the cases we are able to show that our approach significantly improves over standard ant colony optimization and over several competing negative learning mechanisms from the literature.