Conduite dans le but de rendre les robots comme socio-communicatifs, les chercheurs ont cherché à mettre au point des robots dotés de compétences sociales et de «bon sens» pour les rendre acceptables. Cette intelligence sociale ou «sens commun» du robot est ce qui finit par déterminer son acceptabilité sociale à long terme.Cependant, ce n'est pas commun. Les robots peuvent donc seulement apprendre à être acceptables avec l'expérience. Cependant, en enseignant à un humanoïde, les subtilités d'une interaction sociale ne sont pas évidentes. Même un échange de dialogue standard intègre le panel le plus large possible de signes qui interviennent dans la communication et sont difficiles à codifier (synchronisation entre l'expression du corps, le visage, le ton de la voix, etc.). Dans un tel scénario, l'apprentissage du modèle comportemental du robot est une approche prometteuse. Cet apprentissage peut être réalisé avec l'aide de techniques d'IA. Cette étude tente de résoudre le problème de l'apprentissage des modèles comportementaux du robot dans le paradigme automatisé de planification et d'ordonnancement (APS) de l'IA. Dans le domaine de la planification automatisée et de l'ordonnancement (APS), les agents intelligents nécessitent un modèle d'action (plans d'actions dont les exécutions entrelacées effectuent des transitions de l'état système) afin de planifier et résoudre des problèmes réels. Au cours de cette thèse, nous présentons deux nouveaux systèmes d'apprentissage qui facilitent l'apprentissage des modèles d'action et élargissent la portée de ces nouveaux systèmes pour apprendre les modèles de comportement du robot. Ces techniques peuvent être classées dans les catégories non optimale et optimale. Les techniques non optimales sont plus classiques dans le domaine, ont été traitées depuis des années et sont de nature symbolique. Cependant, ils ont leur part de quirks, ce qui entraîne un taux d'apprentissage moins élevé que souhaité. Les techniques optimales sont basées sur les progrès récents dans l'apprentissage en profondeur, en particulier la famille à long terme (LSTM) de réseaux récurrents récurrents. Ces techniques sont de plus en plus séduisantes par la vertu et produisent également des taux d'apprentissage plus élevés. Cette étude met en vedette ces deux techniques susmentionnées qui sont testées sur des repères d'IA pour évaluer leurs prouesses. Ils sont ensuite appliqués aux traces HRI pour estimer la qualité du modèle de comportement du robot savant. Ceci est dans l'intérêt d'un objectif à long terme d'introduire l'autonomie comportementale dans les robots, afin qu'ils puissent communiquer de manière autonome avec les humains sans avoir besoin d'une intervention de «magicien». / Driven with the objective of rendering robots as socio-communicative, there has been a heightened interest towards researching techniques to endow robots with social skills and ``commonsense'' to render them acceptable. This social intelligence or ``commonsense'' of the robot is what eventually determines its social acceptability in the long run.Commonsense, however, is not that common. Robots can, thus, only learn to be acceptable with experience. However, teaching a humanoid the subtleties of a social interaction is not evident. Even a standard dialogue exchange integrates the widest possible panel of signs which intervene in the communication and are difficult to codify (synchronization between the expression of the body, the face, the tone of the voice, etc.). In such a scenario, learning the behavioral model of the robot is a promising approach. This learning can be performed with the help of AI techniques. This study tries to solve the problem of learning robot behavioral models in the Automated Planning and Scheduling (APS) paradigm of AI. In the domain of Automated Planning and Scheduling (APS), intelligent agents by virtue require an action model (blueprints of actions whose interleaved executions effectuates transitions of the system state) in order to plan and solve real world problems. During the course of this thesis, we introduce two new learning systems which facilitate the learning of action models, and extend the scope of these new systems to learn robot behavioral models. These techniques can be classified into the categories of non-optimal and optimal. Non-optimal techniques are more classical in the domain, have been worked upon for years, and are symbolic in nature. However, they have their share of quirks, resulting in a less-than-desired learning rate. The optimal techniques are pivoted on the recent advances in deep learning, in particular the Long Short Term Memory (LSTM) family of recurrent neural networks. These techniques are more cutting edge by virtue, and produce higher learning rates as well. This study brings into the limelight these two aforementioned techniques which are tested on AI benchmarks to evaluate their prowess. They are then applied to HRI traces to estimate the quality of the learnt robot behavioral model. This is in the interest of a long term objective to introduce behavioral autonomy in robots, such that they can communicate autonomously with humans without the need of ``wizard'' intervention.

En informatique ambiante, les objectifs de l'utilisateur peuvent émerger opportunément. Il devient, dès lors, nécessaire de générer à la volée des systèmes interactifs. Un système interactif est composé d'un noyau fonctionnel et d'une Interface Homme-Machine (IHM). Cette thèse traite de la composition d'IHM pour un objectif utilisateur et un contexte d'usage (utilisateur, plate-forme, environnement) donnés. Elle en propose un espace problème fondé sur les exigences utilisateur recueillies par une étude qualitative. Un état de l'art positionne notre travail et en montre la complémentarité par rapport aux travaux existants : la composition du modèle de tâches. La composition de l'IHM concrète est déléguée à une boîte à outils d'interacteurs définis au niveau tâches. La composition du modèle de tâches se fait par planification automatique. L'étude montre que les planificateurs existants répondent partiellement au problème. Aussi, un planificateur a été spécifiquement développé pour l'IHM. Son utilisation est illustrée dans un prototype Compose. Le travail est original à deux titres : d'une part, son approche « Composition de modèles de tâches » est une extension de la littérature ; d'autre part, la composition d'IHM est un nouveau cadre applicatif pour les algorithmes de planification. / In ubiquitous computing, user needs may opportunistically emerge along the variation of the context of use. Thus, there is a need for dynamically composing interactive systems. An interactive system is made of a functional core and a User Interface (UI). This work deals with the composition of UIs to support opportunistic user needs in a given context of use (user, platform, environment). It proposes a problem space of UI composition based on a social study. A state of the art shows the originality of the work: the composition of the task model. The composition of the concrete UI is delegated to a toolkit of interactors defined at the task level. The composition of the task model is done by automated planning. The work shows that current planners do not fulfill Human Computer Interaction (HCI) requirements. Therefore, a specific planner has been developed to compose UIs. This planner is used in Compose, our proof of concept. The work is original in two points: first, by the high level of abstraction the composition is performed at; secondly, by the use of automated planning in HCI.

Planning is a central research area in artificial intelligence, and a lot of effort has gone into constructing more and more efficient planning algorithms. In real-world examples, many problem instances do not have a solution. Hence, there is an obvious need for methods that are capable of identifying unsolvable instances efficiently. It is not possible to efficiently identify all unsolvable instances due to the inherent high complexity of planning, but many unsolvable instances can be identified in polynomial time. We present a number of novel methods for doing this. We adapt the notion of k-consistency (a well-studied concept from constraint satisfaction) for testing unsolvability of planning instances. The idea is to decompose a given problem instance into a number of smaller instances which can be solved in polynomial time. If any of the smaller instances are unsolvable, then the original instance is unsolvable. If all the smaller instances are solvable, then it is possible to extract information which can be used to guide the search. For instance, we introduce the notion of forbidden state patterns that are partial states that must be avoided by any solution to the problem instance. This can be viewed as the opposite of pattern databases which give information about states which can lead to a solution.  We also introduce the notion of critical sets and show how to identify them. Critical sets describe operators or values which must be used or achieved in any solution. It is a variation on the landmark concept, i.e., operators or values which must be used in every solution. With the help of critical sets we can identify superfluous operators and values. These operators and values can be removed by preprocessing the problem instance to decrease planning time.

Les algorithmes permettant la création de stratégies efficaces pour la résolution d’ensemble de problèmeshétéroclites ont toujours été un des piliers de la recherche en Intelligence Artificielle. Dans cette optique,la planification de tâches a pour objectif de fournir à un système la capacité de raisonner pour interagiravec son environnement de façon autonome afin d’atteindre les buts qui lui ont été assignés. À partir d’unedescription de l’état initial du monde, des actions que le système peut exécuter, et des buts qu’il doit atteindre,un planificateur calcule une séquence d’actions dont l’exécution permet de faire passer l’état du monde danslequel évolue le système vers un état qui satisfait les buts qu’on lui a fixés. Le problème de planification esten général difficile à résoudre (PSPACE-difficile), cependant certaines propriétés des problèmes peuvent êtreautomatiquement extraites permettant ainsi une résolution efficace.Dans un premier temps, nous avons développé l’algorithme LMBFS (Landmark-based Meta Best-First Search).À contre-courant des planificateurs state-of-the-art, basés sur la recherche heuristique dans l’espace d’états,LMBFS est un algorithme qui réactualise la technique de décomposition des problèmes de planification baséssur les landmarks. Un landmark est un fluent qui doit être vrai à un certain moment durant l’exécutionde n’importe quel plan solution. L’algorithme LMBFS découpe le problème principal en un ensemble desous-problèmes et essaie de trouver une solution globale grâce aux solutions trouvées pour ces sous-problèmes.Dans un second temps, nous avons adapté un ensemble de techniques pour améliorer les performances del’algorithme. Enfin, nous avons testé et comparé chacune de ces méthodes permettant ainsi la création d’unplanificateur efficace. / The algorithms allowing on-the-fly computation of efficient strategies solving aheterogeneous set of problems has always been one of the greatest challengesfaced by research in Artificial Intelligence. To this end, classical planningprovides to a system reasoning capacities, in order to help it to interact with itsenvironment autonomously. Given a description of the world current state, theactions the system is able to perform, and the goal it is supposed to reach, a plannercan compute an action sequence yielding a state satisfying the predefined goal. Theplanning problem is usually intractable (PSPACE-hard), however some propertiesof the problems can be automatically extracted allowing the design of efficientsolvers.Firstly, we have developed the Landmark-based Meta Best-First Search (LMBFS)algorithm. Unlike state-of-the-art planners, usually based on state-space heuristicsearch, LMBFS reenacts landmark-based planning problem decomposition. Alandmark is a fluent appearing in each and every solution plan. The LMBFSalgorithm splits the global problem in a set of subproblems and tries to find aglobal solution using the solutions found for these subproblems. Secondly, wehave adapted classical planning techniques to enhance the performance of ourbase algorithm, making LMBFS a competitive planner. Finally, we have tested andcompared these methods.

abstract: The rapid advancements of technology have greatly extended the ubiquitous nature of smartphones acting as a gateway to numerous social media applications. This brings an immense convenience to the users of these applications wishing to stay connected to other individuals through sharing their statuses, posting their opinions, experiences, suggestions, etc on online social networks (OSNs). Exploring and analyzing this data has a great potential to enable deep and fine-grained insights into the behavior, emotions, and language of individuals in a society. This proposed dissertation focuses on utilizing these online social footprints to research two main threads – 1) Analysis: to study the behavior of individuals online (content analysis) and 2) Synthesis: to build models that influence the behavior of individuals offline (incomplete action models for decision-making). A large percentage of posts shared online are in an unrestricted natural language format that is meant for human consumption. One of the demanding problems in this context is to leverage and develop approaches to automatically extract important insights from this incessant massive data pool. Efforts in this direction emphasize mining or extracting the wealth of latent information in the data from multiple OSNs independently. The first thread of this dissertation focuses on analytics to investigate the differentiated content-sharing behavior of individuals. The second thread of this dissertation attempts to build decision-making systems using social media data. The results of the proposed dissertation emphasize the importance of considering multiple data types while interpreting the content shared on OSNs. They highlight the unique ways in which the data and the extracted patterns from text-based platforms or visual-based platforms complement and contrast in terms of their content. The proposed research demonstrated that, in many ways, the results obtained by focusing on either only text or only visual elements of content shared online could lead to biased insights. On the other hand, it also shows the power of a sequential set of patterns that have some sort of precedence relationships and collaboration between humans and automated planners. / Dissertation/Thesis / Doctoral Dissertation Computer Science 2019

Planning is the sub-area of artificial intelligence that studies the process of selecting actions to lead an agent, e.g. a robot or a softbot, to a goal state. In many realistic scenarios, any choice of actions can lead the robot into a dead-end state, that is, a state from which the goal cannot be reached. In such cases, the robot can, pro-actively, resort to human help in order to reach the goal, an approach called symbiotic autonomy. In this work, we propose two different approaches to tackle this problem: (I) contingent planning, where the initial state is partially observable, configuring a belief state, and the outcomes of the robot actions are non-deterministic; and (II) probabilistic planning, where the initial state may be partially or totally observable and the actions have probabilistic outcomes. In both approaches, the human help is considered a scarce resource that should be used only when necessary. In contingent planning, the problem is to find a policy (a function mapping belief states into actions) that: (i) guarantees the agent will always reach the goal (strong policy); (ii) guarantees that the agent will eventually reach the goal (strong cyclic policy), or (iii) does not guarantee achieving the goal (weak policy). In this scenario, we propose a contingent planning system that considers human help to transform weak policies into strong (cyclic) policies. To do so, two types of human help are included: (i) human actions that modify states and/or belief states; and (ii) human observations that modify belief states. In probabilistic planning, the problem is to find a policy (a function mapping between world states and actions) that can be one of these two types: a proper policy, where the agent has probability 1 of reaching the goal; or an improper policy, in the case of unavoidable dead-ends. In general, the goal of the agent is to find a policy that minimizes the expected accumulated cost of the actions while maximizes the probability of reaching the goal. In this scenario, this work proposes probabilistic planners that consider human help to transform improper policies into proper policies however, considering two new (alternative) criteria: either to minimize the probability of using human actions or to minimize the expected number of human actions. Furthermore, we show that optimal policies under these criteria can be efficiently computed either by increasing human action costs or given a penalty when a human help is used. Solutions proposed in both scenarios, contingent planning and probabilistic planning with human help, were evaluated over a collection of planning problems with dead-ends. The results show that: (i) all generated policies (strong (cyclic) or proper) include human help only when necessary; and (ii) we were able to find policies for contingent planning problems with up to 10^15000 belief states and for probabilistic planning problems with more than 3*10^18 physical states. / Planejamento é a subárea de Inteligência Artificial que estuda o processo de selecionar ações que levam um agente, por exemplo um robô, de um estado inicial a um estado meta. Em muitos cenários realistas, qualquer escolha de ações pode levar o robô para um estado que é um beco-sem-saída, isto é, um estado a partir do qual a meta não pode ser alcançada. Nestes casos, o robô pode, pró-ativamente, pedir ajuda humana para alcançar a meta, uma abordagem chamada autonomia simbiótica. Neste trabalho, propomos duas abordagens diferentes para tratar este problema: (I) planejamento contingente, em que o estado inicial é parcialmente observável, configurando um estado de crença, e existe não-determinismo nos resultados das ações; e (II) planejamento probabilístico, em que o estado inicial é totalmente observável e as ações tem efeitos probabilísticos. Em ambas abordagens a ajuda humana é considerada um recurso escasso e deve ser usada somente quando estritamente necessária. No planejamento contingente, o problema é encontrar uma política (mapeamento entre estados de crença e ações) com: (i) garantia de alcançar a meta (política forte); (ii) garantia de eventualmente alcançar a meta (política forte-cíclica), ou (iii) sem garantia de alcançar a meta (política fraca). Neste cenário, uma das contribuições deste trabalho é propor sistemas de planejamento contingente que considerem ajuda humana para transformar políticas fracas em políticas fortes (cíclicas). Para isso, incluímos ajuda humana de dois tipos: (i) ações que modificam estados do mundo e/ou estados de crença; e (ii) observações que modificam estados de crenças. Em planejamento probabilístico, o problema é encontrar uma política (mapeamento entre estados do mundo e ações) que pode ser de dois tipos: política própria, na qual o agente tem probabilidade 1 de alcançar a meta; ou política imprópria, caso exista um beco-sem-saída inevitável. O objetivo do agente é, em geral, encontrar uma política que minimize o custo esperado acumulado das ações enquanto maximize a probabilidade de alcançar a meta. Neste cenário, este trabalho propõe sistemas de planejamento probabilístico que considerem ajuda humana para transformar políticas impróprias em políticas próprias, porém considerando dois novos critérios: minimizar a probabilidade de usar ações do humano e minimizar o número esperado de ações do humano. Mostramos ainda que políticas ótimas sob esses novos critérios podem ser computadas de maneira eficiente considerando que ações humanas possuem um custo alto ou penalizando o agente ao pedir ajuda humana. Soluções propostas em ambos cenários, planejamento contingente e planejamento probabilístico com ajuda humana, foram empiricamente avaliadas sobre um conjunto de problemas de planejamento com becos-sem-saida. Os resultados mostram que: (i) todas as políticas geradas (fortes (cíclicas) ou próprias) incluem ajuda humana somente quando necessária; e (ii) foram encontradas políticas para problemas de planejamento contingente com até 10^15000 estados de crença e para problemas de planejamento probabilístico com até 3*10^18 estados do mundo.

Le concept de systèmes informatisés complexes rassemble tous les systèmes constitués d’un grand nombre de composantes inter-connectées et gérées par ordinateur. La configuration et la gestion de ces systèmes passe par une multitude de tâches critiques à leur bon fonctionnement et leur évolution. La problématique de la mise en place et la maîtrise des procédures de gestion et de configuration de tels systèmes devient un point critique au vu de la complexité croissante et du besoin d’évolution de ces derniers. L’objectif de cette thèse est de répondre à cette problématique, à travers le développement d’un outil de planification en mixed-initiative pour la co-génération automatique d’actions de gestion et de configuration, indépendamment du domaine d’application. Dans cette perspective, nous développons « Grounded Totally Ordered Hierarchical Planner » (GTOHP), un planificateur automatique hiérarchique, en « Hierarchical Task Network » (HTN), qui présente des performances élevées nécessaires à une interaction en mixed-initiative. Pour cela nous proposons un algorithme d’instanciation et de simplification des problèmes de planification HTN qui réduit de manière très significative leur complexité et améliore de ce fait les performances des algorithmes de planification. Nous proposons aussi une extension au langage de définition des domaines de planification automatique PDDL afin de modéliser les connaissances des experts du domaine d’application sous forme de méthodes de décomposition des tâches qui serviront à guider l’algorithme de planification HTN. Ensuite, nous intégrons au planificateur GTOHP des mécanismes de récolte de statistiques et d’in- formations sur les résultats des tests réalisés lors de la recherche de plans et nous l’intégrons dans le système « Mixed-Initiative Planner » (MIP) qui fournit plusieurs fonctionnalités d’interaction en mixed-initiative. Nous démontrons les performances élevées du planificateur GTOHP et les apports de l’algorithme d’instanciation et de simplification en le comparant à un planificateur HTN de l’état de l’art à travers une série d’expérimentations sur des domaines de planification issues de la compé- tition internationale de planification automatique. Enfin, nous proposons des critères d’évaluation pour les systèmes en mixed-initiative qui servent de base à la discussion du système MIP. / The concept of complex IT systems includes all systems consisting of a large number of inter-connected and computer-managed components. The configuration and management of these systems involves a multitude of tasks that are critical to their proper functioning and their evolution. The problem of defining procedures for managing and configuring such systems becomes very critical in view of their increasing complexity and their rapid evolution. The aim of this thesis is to develop a mixed-initiative planning tool for the automatic co-generation of a set of management and configuration actions, regardless of the application domain. In this perspective, we develop GTOHP, a hierarchical automatic planner, with HTN, that present the high performance needed for a mixed-initiative interaction. We propose an algorithm for the instantiation and the simplification of HTN planning problems, which significantly reduces their complexity and improves the performance of the planning algorithms. We also propose an extension to the « Planning Domain Definition Language » (PDDL) in order to modelize the knowledge of domain experts in the form of tasks decomposition methods that will be used to guide the HTN planning algorithm. Then, we integrate some mechanisms to GTOHP for collecting statistics and information about the tests results carried out during the plans search and integrate them into the MIP which provides several features of mixed-initiative interaction. We demonstrate the high performance of the GTOHP planner and the contributions of the instantiation and simplification algorithm, by comparing them to a state-of-the-art HTN planner through a series of experiments on planning domains from the international planning competitions. Finally, we propose a panel of evaluation criteria of mixed-initiative systems that serve as a basis for the discussion about the performances and contributions of the MIP system.

Le sujet de recherche est dans la continuité des travaux réalisés au cours de mon M2R sur la programmation par démonstration appliqué à la cobotique en milieu industriel. Ce sujet est à la croisée de plusieurs domaines (interaction Humain-Robot, planification automatique, apprentissage artificiel). Il s'agit maintenant d'aller au delà de ces premiers résultats obtenus au cours de mon M2R et de trouver un cadre générique pour la programmation de « cobots » (robots collaboratifs) en milieu industriel. L'approche cobotique consiste à ce qu'un opérateur humain, en tant qu'expert métier directement impliqué dans la réalisation des tâches en ligne, apprenne au robot à effectuer de nouvelles tâches et à utiliser le robot comme assistant « agile ». Dans ce contexte la thèse propose un mode d'apprentissage de type « end-user programming », c'est-à-dire simple et ne nécessitant pas d'être expert en robotique pour programmer le robot industriel Baxter. / The increasing presence of robots in industries has not gone unnoticed.Cobots (collaborative robots) are revolutionising industries by allowing robots to work in close collaboration with humans.Large industrial players have incorporated them into their production lines, but smaller companies hesitate due to high initial costs and the lack of programming expertise.In this thesis we introduce a framework that combines two disciplines, Programming by Demonstration and Automated Planning, to allow users without programming knowledge to program a robot.The user constructs the robot's knowledge base by teaching it new actions by demonstration, and associates their semantic meaning to enable the robot to reason about them.The robot adopts a goal-oriented behaviour by using automated planning techniques, where users teach action models expressed in a symbolic planning language.In this thesis we present preliminary work on user experiments using a Baxter Research Robot to evaluate our approach.We conducted qualitative user experiments to evaluate the user's understanding of the symbolic planning language and the usability of the framework's programming process.We showed that users with little to no programming experience can adopt the symbolic planning language, and use the framework.We further present our work on a Programming by Demonstration system used for organisation tasks.The system includes a goal inference model to accelerate the programming process by predicting the user's intended product configuration.

No description available.

Artificial Intelligence Planning is about acting in order to achieve a desired goal. Under incomplete information, the task of finding the actions needed to achieve the goal can be modelled as a search problem in the belief space. This task is costly, as belief space is exponential in the number of states, which is exponential in the number of variables. Good belief representations and heuristics are thus critical for scaling up in this setting. The translation-based approach to automated planning with incomplete information deals with both issues by casting the problem of search in belief space to a search problem in state space, where each node of the search space represents a belief state. We develop plan synthesis tools that use translated versions of planning problems under uncertainty, with partial or null sensing available. We show formally under which conditions the introduced translations are polynomial, and capture all and only the plans of the original problems. We study empirically the value of these translations. / La Planificación es la disciplina de Inteligencia Artificial que estudia los procesos de razonamiento necesarios para conseguir las acciones que logren un objetivo dado. En presencia de información incompleta, el problema de planificación puede ser modelado como una búsqueda en el espacio de estados de creencia, cada uno de ellos representando un conjunto de estados posibles. Este problema es costoso ya que el numero de estados de creencia puede ser exponencial en el número de estados, lo cual es exponencial en el número de variables del problema. El uso de buenas representaciónes de los estados y de heurísticas informadas resultan cruciales para escalar en este espacio de búsqueda. En esta tesis se presentan traducciones para planificación con información incompleta, que transforman el problema de búsqueda en el espacio de estados de creencia, en búsqueda en espacio de estados, donde cada nodo representa un estado de creencia. Hemos desarrollado herramientas para la generación de planes para el problema traducido, ya sea con percepción parcial o nula. A su vez, demostramos formalmente bajo qué circunstancias las traducciones son polinómicas, completas y correctas. La evaluación empírica remarca el valor de dichas traducciones