This master thesis investigates the principal capability of artificial evolution to produce tool use behavior in adaptive agents, excluding the application of life-time learning or adaptation mechanisms. Tool use is one aspect of complex behavior that is expected from autonomous agents acting in real-world environments. In order to achieve tool use behavior an agent needs to identify environmental objects as potential tools before it can use the tools in a problem-solving task. Up to now research in robotics has focused on life-time learning mechanisms in order to achieve this. However, these techniques impose great demands on resources, e.g. in terms of memory or computational power. All of them have shown limited results with respect to a general adaptivity. One might argue that even nature does not present any kind of omni-adaptive agent. While humans seem to be a good example of natural agents that master an impressive variety of life conditions and environments (at least from a human perspective, other examples are spectacular survivability observations of octopuses, scorpions or various viruses) even the most advanced engineering approaches can hardly compete with the simplest life-forms in terms of adaptation. This thesis tries to contribute to engineering approaches by promoting the application of artificial evolution as a complementing element with the presentation of successful pioneering experiments. The results of these experiments show that artificial evolution is indeed capable to render tool use behavior at different levels of complexity and shows that the application of artificial evolution might be a good complement to life-time approaches in order to create agents that are able to implicitly extract concepts and display tool use behavior. The author believes that off-loading at least parts of the concept retrieval process to artificial evolution will reduce resource efforts at life-time when creating autonomous agents with complex behavior such as tool use. This might be a first step towards the vision of a higher level of autonomy and adaptability. Moreover, it shows the demand for an experimental verification of commonly accepted limits between qualities of learned and evolved tool use capabilities.

Custom software development and maintenance is one of the key expenses associated with developing automated systems for mass customization. This paper presents a method for reducing the risk associated with this expense by developing a flexible environment for determining and executing dynamic workflow paths. Strategies for developing an autonomous agent-based framework and for identifying and creating web services for specific process tasks are presented. The proposed methods are outlined in two different case studies to illustrate the approach for both a generic process with complex workflow paths and a more specific sequential engineering process.

In the area of steering behaviours of autonomous agents and crowd simulations, there is a plethora of methods for executing the simulations. A very hard-to-achieve goal of crowd simulations is to make them seem natural and accurately reflect real-life crowds. A very important criterion for this goal is to have the agents avoid collisions, both with each other and with the environment. A less important, but important nonetheless, criterion is to not let the time taken or distance covered to reach the goal in the simulation be too high, compared with when not implementing collision avoidance. This paper proposes and explores a novel method of enhancing vector field-based steering with rule-based deviations to implement collision avoidance. This method is called ’DevVec’ (’Deviation + Vector Field steering’). The rules which are used for the deviations are extracted from a user survey, and they describe what the agent should do in different collision avoidance scenarios. The viability of DevVec is tested by comparing it with another already established method, called ’Gradient-based Steering’, in terms of fulfilling the criteria mentioned above. Both methods are used to simulate pedestrians moving throughout different scenes. The results suggest that DevVec has potential, but would require additional time and resources, and perhaps a few changes in future works to be presented in its best possible version. / Inom ämnesområdet för styrbeteenden hos autonoma agenter och simuleringar av folkmassor finns det många metoder för att framställa dessa simuleringar. Ett väldigt svåruppnåeligt mål för denna typ av simuleringar är få dem att verka naturliga och verklighetstrogna. Ett viktigt kriterie för detta mål är att få agenterna att undvika kollisioner, både med varandra och med den kringliggande omgivningen. Ett mindre viktigt, men viktigt oavsett, kriterie är att inte låta en agent ta för lång tid eller gå för långt för att nå sitt mål i simuleringen, i jämförelse med när de inte försöka undvika hinder. Denna studie presenterar och utforskar en ny metod som utökar en vektorfältsbaserat styralgoritm med regelbaserade avvikelser för att ta hänsyn till att undvika kollisioner. Denna nya metod kallas för ’DevVec’ (’Deviation + Vector Field steering’). Reglerna som används för avvikelserna är framtagna från en enkät, och de beskriver vad en agent borde göra vid olika kollision-scenarion. Användbarheten av DevVec prövas genom att jämföra den med en redan etablerad metod som kallas för ’Gradientbaserad styrning’, med avseende på de ovan nämnda kriterierna. Båda metoderna används för att simulera fotgängare i olika omgivningar. Resultaten antyder att DevVec har potential, men att det krävs ytterligare tid och resurser, och troligtvis några ändringar i framtiden för att framställa den bästa möjliga versionen.

This thesis addresses the problem of positional consensus of multi-agent systems. A positional consensus is achieved when the agents converge to a point. Some applications of this class of problem is in mid-air refueling of the aircraft or UAVs, targeting a geographical location, etc. In this research work some positional consensus algorithms have been developed. They can be categorized in two part (i) Broadcast control based algorithm (ii) Distributed control based algorithm. In case of broadcast based algorithm control strategies for a group of agents is developed to achieve positional consensus. The problem is constrained by the requirement that every agent must be given the same control input through a broadcast communication mechanism. Although the control command is computed using state information in a global framework, the control input is implemented by the agents in a local coordinate frame. The mathematical formulation has been done in a linear programming framework that is computationally less intensive than earlier proposed methods. Moreover, a random perturbation input in the control command, that helps to achieve reasonable proximity among agents even for a large number of agents, which was not possible with the existing strategy in the literature, is introduced. This method is extended to achieve positional consensus at a pre-specified location. A comparison between the LP approach and the existing SOCP based approach is also presented. Some of the algorithm has been demonstrated successfully on a robotic platform made from LEGO Mindstorms NXT Robots. In the second case of broadcast based algorithm, a decentralized algorithm for a group of multiple autonomous agents to achieve positional consensus has been developed using the broadcast concept. Even here, the mathematical formulation has done using a linear programming framework. Each agent has some sensing radius and it is capable of sensing position and orientation with other agents within their sensing region. The method is computationally feasible and easy to implement. In case of distributed algorithms, a computationally efficient distributed rendezvous algorithm for a group of autonomous agents has been developed. The algorithm uses a rectilinear decision domain (RDD), as against the circular decision domain assumed in earlier work available in the literature. This helps in reducing its computational complexity considerably. An extensive mathematical analysis has been carried out to prove the convergence of the algorithm. The algorithm has also been demonstrated successfully on a robotic platform made from LEGO Mindstorms NXT Robots.

Les robots opèrent dans le monde réel, dans lequel essayer quelque chose prend beaucoup de temps. Pourtant, les methodes d’apprentissage par renforcement actuels (par exemple, deep reinforcement learning) nécessitent de longues périodes d’interaction pour trouver des politiques efficaces. Dans cette thèse, nous avons exploré des algorithmes qui abordent le défi de l’apprentissage par essai-erreur en quelques minutes sur des robots physiques. Nous appelons ce défi “Apprentissage par renforcement micro-data”. Dans la première contribution, nous avons proposé un nouvel algorithme d’apprentissage appelé “Reset-free Trial-and-Error” qui permet aux robots complexes de s’adapter rapidement dans des circonstances inconnues (par exemple, des dommages) tout en accomplissant leurs tâches; en particulier, un robot hexapode endommagé a retrouvé la plupart de ses capacités de marche dans un environnement avec des obstacles, et sans aucune intervention humaine. Dans la deuxième contribution, nous avons proposé un nouvel algorithme de recherche de politique “basé modèle”, appelé Black-DROPS, qui: (1) n’impose aucune contrainte à la fonction de récompense ou à la politique, (2) est aussi efficace que les algorithmes de l’état de l’art, et (3) est aussi rapide que les approches analytiques lorsque plusieurs processeurs sont disponibles. Nous avons aussi proposé Multi-DEX, une extension qui s’inspire de l’algorithme “Novelty Search” et permet de résoudre plusieurs scénarios où les récompenses sont rares. Dans la troisième contribution, nous avons introduit une nouvelle procédure d’apprentissage du modèle dans Black-DROPS qui exploite un simulateur paramétré pour permettre d’apprendre des politiques sur des systèmes avec des espaces d’état de grande taille; par exemple, cette extension a trouvé des politiques performantes pour un robot hexapode (espace d’état 48D et d’action 18D) en moins d’une minute d’interaction. Enfin, nous avons exploré comment intégrer les contraintes de sécurité, améliorer la robustesse et tirer parti des multiple a priori en optimisation bayésienne. L'objectif de la thèse était de concevoir des méthodes qui fonctionnent sur des robots physiques (pas seulement en simulation). Par conséquent, tous nos approches ont été évaluées sur au moins un robot physique. Dans l’ensemble, nous proposons des méthodes qui permettre aux robots d’être plus autonomes et de pouvoir apprendre en poignée d’essais / Robots have to face the real world, in which trying something might take seconds, hours, or even days. Unfortunately, the current state-of-the-art reinforcement learning algorithms (e.g., deep reinforcement learning) require big interaction times to find effective policies. In this thesis, we explored approaches that tackle the challenge of learning by trial-and-error in a few minutes on physical robots. We call this challenge “micro-data reinforcement learning”. In our first contribution, we introduced a novel learning algorithm called “Reset-free Trial-and-Error” that allows complex robots to quickly recover from unknown circumstances (e.g., damages or different terrain) while completing their tasks and taking the environment into account; in particular, a physical damaged hexapod robot recovered most of its locomotion abilities in an environment with obstacles, and without any human intervention. In our second contribution, we introduced a novel model-based reinforcement learning algorithm, called Black-DROPS that: (1) does not impose any constraint on the reward function or the policy (they are treated as black-boxes), (2) is as data-efficient as the state-of-the-art algorithm for data-efficient RL in robotics, and (3) is as fast (or faster) than analytical approaches when several cores are available. We additionally proposed Multi-DEX, a model-based policy search approach, that takes inspiration from novelty-based ideas and effectively solved several sparse reward scenarios. In our third contribution, we introduced a new model learning procedure in Black-DROPS (we call it GP-MI) that leverages parameterized black-box priors to scale up to high-dimensional systems; for instance, it found high-performing walking policies for a physical damaged hexapod robot (48D state and 18D action space) in less than 1 minute of interaction time. Finally, in the last part of the thesis, we explored a few ideas on how to incorporate safety constraints, robustness and leverage multiple priors in Bayesian optimization in order to tackle the micro-data reinforcement learning challenge. Throughout this thesis, our goal was to design algorithms that work on physical robots, and not only in simulation. Consequently, all the proposed approaches have been evaluated on at least one physical robot. Overall, this thesis aimed at providing methods and algorithms that will allow physical robots to be more autonomous and be able to learn in a handful of trials

<p>Autonomous systems are often deployed in dynamic environments and are challenged with unexpected changes (novelties) in the environments where they receive novel data that was not seen during training. Given the uncertainty, they should be able to operate without (or with limited) human intervention and they are expected to (1) Adapt to such changes while still being effective and efficient in performing their multiple tasks. The system should be able to provide continuous availability of its critical functionalities. (2) Make informed decisions independently from any central authority. (3) Be Cognitive: learns the new context, its possible actions, and be rich in knowledge discovery through mining and pattern recognition. (4) Be Reflexive: reacts to novel unknown data as well as to security threats without terminating on-going critical missions. These characteristics combine to create the workflow of autonomous decision-making process in multi-agent environments (i.e.,) any action taken by the system must go through these characteristic models to autonomously make an ideal decision based on the situation. </p> <p><br></p> <p>In this dissertation, we propose novel learning-based models to enhance the decision-making process in autonomous multi-agent systems where agents are able to detect novelties (i.e., unexpected changes in the environment), and adapt to it in a timely manner. For this purpose, we explore two complex and highly dynamic domains </p> <p>(1) Transportation Networks (e.g., Ridesharing application): where we develop AdaPool: a novel distributed diurnal-adaptive decision-making framework for multi-agent autonomous vehicles using model-free deep reinforcement learning and change point detection. (2) Multi-agent games (e.g., Monopoly): for which we propose a hybrid approach that combines deep reinforcement learning (for frequent but complex decisions) with a fixed-policy approach (for infrequent but straightforward decisions) to facilitate decision-making and it is also adaptive to novelties. (3) Further, we present a domain agnostic approach for decision making without prior knowledge in dynamic environments using Bootstrapped DQN. Finally, to enhance security of autonomous multi-agent systems, (4) we develop a machine learning based resilience testing of address randomization moving target defense. Additionally, to further  improve the decision-making process, we present (5) a novel framework for multi-agent deep covering option discovery that is designed to accelerate exploration (which is the first step of decision-making for autonomous agents), by identifying potential collaborative agents and encouraging visiting the under-represented states in their joint observation space. </p>

Behavioral animation has become popular for creating virtual characters that are autonomous agents and thus self-animating. This is useful for lessening the workload of human animators, populating virtual environments with interactive agents, etc. Unfortunately, current behavioral animation techniques suffer from three key problems: (1) deliberative behavioral models (i.e., cognitive models) are slow to execute; (2) interactive virtual characters cannot adapt online due to interaction with a human user; (3) programming of behavioral models is a difficult and time-intensive process. This dissertation presents a collection of papers that seek to overcome each of these problems. Specifically, these issues are alleviated through novel machine learning schemes. Problem 1 is addressed by using fast regression techniques to quickly approximate a cognitive model. Problem 2 is addressed by a novel multi-level technique composed of custom machine learning methods to gather salient knowledge with which to guide decision making. Finally, Problem 3 is addressed through programming-by-demonstration, allowing a non technical user to quickly and intuitively specify agent behavior.

Cognitive and behavioral models have become popular methods to create autonomous self-animating characters. Creating these models presents the following challenges: (1) Creating a cognitive or behavioral model is a time intensive and complex process that must be done by an expert programmer (2) The models are created to solve a specific problem in a given environment and because of their specific nature cannot be easily reused. Combining existing models together would allow an animator, without the need of a programmer, to create new characters in less time and would be able to leverage each model's strengths to increase the character's performance, and to create new behaviors and animations. This thesis provides a framework that can aggregate together existing behavioral and cognitive models into an ensemble. An animator only has to rate how appropriately a character performed and through machine learning the system is able to determine how the character should act given the current situation. Empirical results from multiple case studies validate the approach taken.

<p dir="ltr">Current artificial intelligence (AI) systems are proficient at tasks in a closed-world setting where the rules are often rigid. However, in real-world applications, the environment is usually open and dynamic. In this work, we investigate the effects of such dynamic environments on AI systems and develop ways to mitigate those effects. Central to our exploration is the concept of \textit{novelties}. Novelties encompass structural changes, unanticipated events, and environmental shifts that can confound traditional AI systems. We categorize novelties based on their representation, anticipation, and impact on agents, laying the groundwork for systematic detection and adaptation strategies. We explore novelties in the context of stochastic games. Decision-making in stochastic games exercises many aspects of the same reasoning capabilities needed by AI agents acting in the real world. A multi-agent stochastic game allows for infinitely many ways to introduce novelty. We propose an extension of the deep reinforcement learning (DRL) paradigm to develop agents that can detect and adapt to novelties in these environments. To address the sample efficiency challenge in DRL, we introduce a hybrid approach that combines fixed-policy methods with traditional DRL techniques, offering enhanced performance in complex decision-making tasks. We present a novel method for detecting anticipated novelties in multi-agent games, leveraging information theory to discern patterns indicative of collusion among players. Finally, we introduce DABLER, a pioneering deep reinforcement learning architecture that dynamically adapts to changing environmental conditions through broad learning approaches and environment recognition. Our findings underscore the importance of developing AI systems equipped to navigate the uncertainties of the open world, offering promising pathways for advancing AI research and application in real-world settings.</p>

Cette recherche se caractérise, premièrement, par une discussion théorique sur le concept d'agent autonome, basée sur des éléments issus des paradigmes de l'Intelligence Artificielle Située et de l'Intelligence Artificielle Affective. Ensuite, cette thèse présente le problème de l'apprentissage de modèles du monde, en passant en revue la littérature concernant les travaux qui s'y rapportent. A partir de ces discussions, l'architecture CAES et le mécanisme CALM sont présentes. CAES (Coupled Agent-Environment System) constitue une architecture pour décrire des systèmes bases sur la dichotomie agent-environnement. Il définit l'agent et l'environnement comme deux systèmes partiellement ouverts, en couplage dynamique. Dans CAES, l'agent est compose de deux sous-systèmes, l'esprit et le corps, suivant les principes de la situativite et de la motivation intrinsèque. CALM (Constructivist Anticipatory Learning Mechanism) est un mécanisme d'apprentissage fonde sur l'approche constructiviste de l'Intelligence Artificielle. Il permet a un agent situe de construire un modèle du monde dans des environnements partiellement observables et partiellement déterministes, sous la forme d'un processus de décision markovien partiellement observable et factorise (FPOMDP). Le modèle du monde construit est ensuite utilise pour que l'agent puisse définir une politique d'action visant à améliorer sa propre performance / This research is characterized, first, by a theoretical discussion on the concept of autonomous agent, based on elements taken from the Situated AI and the Affective AI paradigms. Secondly, this thesis presents the problem of learning world models, providing a bibliographic review regarding some related works. From these discussions, the CAES architecture and the CALM mechanism are presented. The CAES (Coupled Agent-Environment System) is an architecture for describing systems based on the agent-environment dichotomy. It defines the agent and the environment as two partially open systems, in dynamic coupling. In CAES, the agent is composed of two sub-systems, mind and body, following the principles of situativity and intrinsic motivation. CALM (Constructivist Learning Anticipatory Mechanism) is based on the constructivist approach to Artificial Intelligence. It allows a situated agent to build a model of the world in environments partially deterministic and partially observable in the form of Partially Observable and Factored Markov Decision Process (FPOMDP). The model of the world is constructed and used for the agent to define a policy for action in order to improve its own performance