In the last few years, as processing the data became a part of everyday life in different areas of human activity, the automated machine learning systems that are designed to help with the process of data mining, are on the rise. Various metalearning techniques, including recommendation of the right method to use, or the sequence of steps to take, and to find its optimum hyperparameters configuration, are integrated into these systems to help the researchers with the machine learning tasks. In this thesis, we proposed metalearning algorithms and techniques for hyperparameters optimization, narrowing the intervals of hyperparameters, and recommendations of a machine learning method for a never before seen dataset. We designed two AutoML machine learning systems, where these metalearning techniques are implemented. The extensive set of experiments was proposed to evaluate these algorithms, and the results are presented.

Language is critical to establishing long-term cooperative relationships among intelligent agents (including people), particularly when the agents' preferences are in conflict. In such scenarios, an agent uses speech to coordinate and negotiate behavior with its partner(s). While recent work has shown that neural language modeling can produce effective speech agents, such algorithms typically only accept previous text as input. However, in relationships among intelligent agents, not all relevant context is expressed in conversation. Thus, in this paper, we propose and analyze an algorithm, called Llumi, that incorporates other forms of context to learn to speak in long-term relationships modeled as repeated games with cheap talk. Llumi combines models of intentionality with neural language modeling techniques to learn speech from data that is relevant to the agent's current context. A user study illustrates that, while imperfect, Llumi does learn context-aware speech repeated games with cheap talk when partnered with people, including games in which it was not trained. We believe these results are useful in determining how autonomous agents can learn to use speech to facilitate successful human-agent teaming.

The problem of determining who should do what given a set of tasks and a set of agents is called the task allocation problem. The problem occurs in many multi-agent system applications where a workload of tasks should be shared by a number of agents. In our case, the task allocation problem occurs as an integral part of a larger problem of determining if a task can be delegated from one agent to another. Delegation is the act of handing over the responsibility for something to someone. Previously, a theory for delegation including a delegation speech act has been specified. The speech act specifies the preconditions that must be fulfilled before the delegation can be carried out, and the postconditions that will be true afterward. To actually use the speech act in a multi-agent system, there must be a practical way of determining if the preconditions are true. This can be done by a process that includes solving a complex task allocation problem by the agents involved in the delegation. In this thesis a constraint-based task specification formalism, a complex task allocation algorithm for allocating tasks to unmanned aerial vehicles and a generic collaborative system shell for robotic systems are developed. The three components are used as the basis for a collaborative unmanned aircraft system that uses delegation for distributing and coordinating the agents' execution of complex tasks.

Nous abordons dans cette thèse un problème important en simulation multi-agent pour l'étude des systèmes complexes: celui d'assembler de multiples expertises par une approche multi-niveau. Alors que les approches existantes considèrent habituellement la vue d'un seul expert principal sur le système, nous proposons d'utiliser une approche multi-niveau pour intégrer plusieurs expertises sous la forme d'agents de différents niveaux d'abstraction. Nous montrons qu'il est ainsi possible de rester proche des concepts manipulés par les différents experts (ce qui permet de faciliter le processus de validation dans leurs domaines respectifs) et de combiner les différents niveaux de ces concepts, de manière à ce que chaque expert puisse comprendre les dynamiques des éléments liés à son domaine. Nous proposons le méta-modèle SIMLAB basé sur une représentation unifiée des concepts par des agents pouvant s'influencer les uns les autres dans différents axes et différents niveaux. Ce travail est concrétisé dans le cadre de l'étude de l'activité humaine en relation avec la consommation électrique. Il s'agit là d'un exemple typique de système complexe nécessitant de multiples expertises issues de différents domaines tels que l'ergonomie, l'énergétique, la sociologie, la thermique, ... Dans ce contexte, nous présentons ensuite la mise en oeuvre de notre approche dans la plate-forme SMACH de simulation des comportements humains et nous décrivons un ensemble d'expérimentations illustrant les différentes caractéristiques de notre approche. Nous montrons enfin la capacité de SIMLAB à reproduire et à étendre en simulation une étude réalisée sur le terrain de gestion de la demande énergétique. / The purpose of this work is to tackle a key problem in the study of complex systems when using multi-agent simulation: how to assemble several domains of expertise with a multi-level approach. While existing approaches usually consider the viewpoint of a unique main expert, we propose to use a multi-level model to integrate the multiple domains of expertise embodied in agents located at different abstraction levels. In this work, we show that it is possible to both stay close to the concepts manipulated by the experts (for the sake of the validation process in the domain of each expert) and combine the levels of those concepts. That way, each expert can easily understand the dynamics of the components related to their domain.We present SIMLAB, our meta-model based on a unified representation of the concepts using agents. Each agent can influence the others on different axes and levels. This work is materialised in a study of human activity in relation to electrical consumption. It is a typical example of complex system which requires many domains of expertise such as psychology, energetics, sociology, heat science, … In this context, we present the implementation of our approach in SMACH, a simulation platform of human behaviours. We Then describe several experiments to illustrate the characteristics of our approach. Finally, we show how SIMLAB can reproduce and extend in silico a field study of energy demand management.

This compilation thesis presents an overarching framework on the utilization of nonlinear model predictive control(NMPC) for various applications in the context of Unmanned Aerial Vehicle (UAV) path planning and collision avoidance. Fast and novel optimization algorithms allow for NMPC formulations with high runtime requirement, as those posed by controlling UAVs, to also have sufficiently large prediction horizons as to in an efficient manner integrate collision avoidance in the form of set-exclusion constraints that constrain the available position-space of the robot. This allows for an elegant merging of set-point reference tracking with the collision avoidance problem, all integrated in the control layer of the UAV. The works included in this thesis presents the UAV modeling, cost functions, constraint definitions, as well as the utilized optimization framework. Additional contributions include the use case on multi-agent systems, how to classify and predict trajectories of moving (dynamic) obstacles, as well as obstacle prioritization when an aerial agent is in the precense of more obstacles, or other aerial agents, than can reasonably be defined in the NMPC formulation. For the cases of dynamic obstacles and for multi-agent distributed collision avoidance this thesis offers extensive experimental validation of the overall NMPC framework. These works push the limits of the State-of-the-Art regarding real-time real-life implementations of NMPC-based collision avoidance. The works also include a novel RRT-based exploration framework that combines path planning with exploration behavior. Here, a multi-path RRT * planner plans paths to multiple pseudo-random goals based on a sensor model and evaluates them based on the potential information gain, distance travelled, and the optimimal actuation along the paths.The actuation is solved for as as the solutions to a NMPC problem, implying that the nonlinear actuator-based and dynamically constrained UAV model is considered as part of the combined exploration plus path planning problem. To the authors best knowledge, this is the first time the optimal actuation has been considered in such a planning problem. For all of these applications, the utilized optimization framework is the Optimization Engine: a code-generation framework that generates a custom Rust-based solver from a specified model, cost function, and constraints. The Optimization Engine solves general nonlinear and nonconvex optimization problems, and in this thesis we offer extensive experimental validation of the utilized Proximal-Averaged Newton-type method for Optimal Control (PANOC) algorithm as well as both the integrated Penalty Method and Augmented Lagrangian Method for handling the nonlinear nonconvex constraints that result from collision avoidance problems.

京都大学 / 0048 / 新制・課程博士 / 博士(工学) / 甲第18573号 / 工博第3934号 / 新制||工||1604(附属図書館) / 31473 / 京都大学大学院工学研究科都市社会工学専攻 / (主査)教授 谷口 栄一, 准教授 宇野 伸宏, 准教授 QURESHIAli Gul / 学位規則第4条第1項該当 / Doctor of Philosophy (Engineering) / Kyoto University / DFAM

This thesis studies the control algorithms used by a team of ECAVs (Electronic Combat Air Vehicle) to deceive a network of radars to detect a phantom track. Each ECAV has the electronic capability of intercepting the radar waves, and introducing an appropriate time delay before transmitting it back, and deceiving the radar into seeing a spurious target beyond its actual position. On the other hand, to avoid the errors and increase the reliability, have a complete coverage in various atmosphere conditions, and confronting the effort of the belligerent intruders to delude the sentinel and enter the area usually a network of radars are deployed to guard the region. However, a team of cooperating ECAVs could exploit this arrangement and plans their trajectories in a way all the radars in the network vouch for seeing a single and coherent spurious track of a phantom. Since each station in the network confirms the other, the phantom track is considered valid. This problem serves as a motivating example in trajectory planning for the multi-agent system in highly constrained operation conditions. The given control command to each agent should be a viable one in the agent limited capabilities, and also drives it in a cumulative action to keep the formation. In this thesis, three different approaches to devise a trajectory for each agent is studied, and the difficulties for deploying each one are addressed. In the first one, a command center has all information about the state of the agents, and in every step decides about the control each agent should apply. This method is very effective and robust, but needs a reliable communication. In the second method, each agent decides on its own control, and the members of the group just communicate and agree on the range of control they like to apply on the phantom. Although in this method much less data needs to communicate between the agents, it is very sensitive to the disturbances and miscalculations, and could be easily fell apart or come to a state with no feasible solution to continue. In the third method a differential geometric approach to the problem is studied. This method has a very strong backbone, and minimizes the communication needed to a binary one. However, less data provided to the agents about the system, more sensitive and infirm the system is when it faced with imperfectionalities. In this thesis, an object oriented program is developed in the Matlab software area to simulate all these three control strategies in a scalable fashion. Object oriented programming is a naturally suitable method to simulate a multi-agent system. It gives the flexibility to make the code more iv close to a real scenario with defining each agent as a separated and independent identity. The main objective is to understand the nature of the constrained dynamic problems, and examine various solutions in different situations. Using the flexibility of this code, we could simulate several scenarios, and incorporate various conditions on the system. Also, we could have a close look at each agent to observe its behavior in these situations. In this way we will gain a good insight of the system which could be used in designing of the agents for specific missions.

In this thesis, new algorithms for formation control of multi agent systems (MAS) based on continuum mechanics principles will be investigated. For this purpose agents of the MAS are treated as particles in a continuum, evolving in an n-D space, whose desired configuration is required to satisfy an admissible deformation function. Considered is a specific class of mappings that is called homogenous where the Jacobian of the mapping is only a function of time and is not spatially varying. The primary objectives of this thesis are to develop the necessary theory and its validation via simulation on a mobile-agent based swarm test bed that includes two primary tasks: 1) homogenous transformation of MAS and 2) deployment of a random distribution of agents on to a desired configuration. Developed will be a framework based on homogenous transformations for the evolution of a MAS in an n-D space (n=1, 2, and 3), under two scenarios: 1) no inter-agent communication (predefined motion plan); and 2) local inter-agent communication. Additionally, homogenous transformations based on communication protocols will be used to deploy an arbitrary distribution of a MAS on to a desired curve. Homogenous transformation with no communication: A homogenous transformation of a MAS, evolving in an space, under zero inter agent communication is first considered. Here the homogenous mapping, is characterized by an n x n Jacobian matrix ( ) and an n x 1 rigid body displacement vector ( ), that are based on positions of n+1 agents of the MAS, called leader agents. The designed Jacobian ( ) and rigid body displacement vector ( ) are passed onto rest of the agents of the MAS, called followers, who will then use that information to update their positions under a pre- iv defined motion plan. Consequently, the motion of MAS will evolve as a homogenous transformation of the initial configuration without explicit communication among agents. Homogenous Transformation under Local Communication: We develop a framework for homogenous transformation of MAS, evolving in , under a local inter agent communication topology. Here we assume that some agents are the leaders, that are transformed homogenously in an n-D space. In addition, every follower agent of the MAS communicates with some local agents to update its position, in order to grasp the homogenous mapping that is prescribed by the leader agents. We show that some distance ratios that are assigned based on initial formation, if preserved, lead to asymptotic convergence of the initial formation to a final formation under a homogenous mapping. Deployment of a Random Distribution on a Desired Manifold: Deployment of agents of a MAS, moving in a plane, on to a desired curve, is a task that is considered as an application of the proposed approach. In particular, a 2-D MAS evolution problem is considered as two 1-D MAS evolution problems, where x or y coordinates of the position of all agents are modeled as points confined to move on a straight line. Then, for every coordinate of MAS evolution, bulk motion is controlled by two agents considered leaders that move independently, with rest of the follower agents motions evolving through each follower agent communicating with two adjacent agents.

In the past few years, the application of Control Barrier Functions (CBF) and High Order Control Barrier Functions (HOCBF) as a suitable framework to ensure safety for autonomous systems has attracted increasing interest. In particular, autonomous space systems are frequently subject to safety-critical constraints due to the high costs involved in manufacturing and launching. In the present work, the application of a sample data MPC controller subject to CBF and HOCBF constraints is explored as a suitable solution for spacecraft formation flight operations. Specifically, a prototype inspection mission of the International Space Station through a multi-agent formation of CubeSats is explored. Each CubeSat is assumed to be injected in a passive relative orbit around the ISS and controlled such that the state of each agent is maintained within a prescribed safe corridor from its reference relative orbit. Moreover, appropriate conditions on the minimum control authority required to guarantee the constraints satisfaction within the MPC scheme formulation are derived and a numerical procedure to assess the recursive feasibility of the designed controller is presented. Moreover, suitable analytical modifications of the classical CBF and HOCBF constraints definitions are introduced such that the presented sample data MPC control scheme is guaranteed to ensure safety for the state of each agent in between sampling intervals. Lastly, the final control strategy is validated numerically by means of computer simulation. / Under de senaste åren har tillämpningen av Kontrollbarriärfunktioner (CBF) och Högre ordningens kontrollbarriärfunktioner (HOCBF) som ett lämpligt ramverk för att säkerställa säkerhet för autonoma system väckt ett ökande intresse. Autonoma rymdsystem är ett område med särskilt fokus på säkerhetsbegränsningar på grund av de höga tillverknings och uppskjutningskostnaderna. I detta arbete undersöks tillämpningen av en MPC-kontroller med CBF och HOCBF bivillkor för applikation inom formationsflygningsoperationer för rymdfarkoster. Detta görs genom att ett prototypinspektionsuppdrag på Internationella Rymdstationen (ISS) genom en multi-agent formation av CubeSats tas fram. Varje CubeSat är ämnad att injiceras i en passiv relativ omloppsbana runt ISS och styras sådant att varje agents tillstånd bevaras inom en föreskriven säker korridor från dess passiva relativa referensomloppsbana. Lämpliga villkor för den minsta styrbarheten som krävs för att garantera att MPC-schemaformuleringens begränsningar är tillfredsställda härleds, och en numerisk procedur för att bedöma den rekursiva genomförbarheten för den designade kontrollern presenteras. Vidare introduceras lämpliga analytiska modifieringar av de klassiska CBF- och HOCBF-begränsningsdefinitionerna så att det presenterade MPC-kontrollschemat med provdata garanterar säkerheten för varje agents tillstånd mellan dess samplingsintervall. Till sist valideras den slutliga kontrollstrategin numeriskt via datorsimuleringar.

This report presents a strategy to control multiple collaborative intelligent agents acting in a complex, versatile environment. The proposed method utilizes back-chained behavior trees and 1-to-1 task distribution. The agents claim a task, which prevents other agents in the system to start working on the same task. Backward chaining is an algorithm for generating reactive agents from a set of goals. The method was evaluated in Minecraft with Microsoft’s Project Malmo API. Two different scenarios were considered. In the first one, a group of agents collaborated to build a structure. In the second one, a group of agents collaborated while gathering material. We propose and evaluate three algorithms with different levels of agent-cooperation and complexity (Algorithm 1, Algorithm 2, and Algorithm 3). The evaluation shows that backward chained Behaviour Trees (BTs) works well for multiagent coordination in complex versatile environments and that adding 1-to-1 task distribution increases the efficiency of the agents when completing the experiment tasks. / Rapporten presenterar en metod för styrning av en grupp kollaborativa intelligenta agenter agerande i en komplex dynamisk miljö. Den förslagna metoden använder sig av bakåtkedjade beteendeträd och 1-mot-1 uppgiftsdistribution, där en agent reserverar en uppgift vilket hindrar andra agenter att börja arbeta på samma uppgift. Bakåtkedjning är en metod som möjliggör generering av flexibla agenter utifrån en lista av mål och krav. Metoden utvärderades i två olika scenarion i tv-spelet Minecraft. Agenterna samarbetar i det första scenariot med att bygga en struktur och i det andra scenariot med att samla material. Vi föreslår och utvärderar tre algoritmer med olika nivåer av agentsamarbete och komplexitet (Algoritm 1, Algoritm 2, och Algorithm 3). Utvärderingerarna indikerar att bakåtkedjade beteendeträd fungerar bra för multiagentkoordination i komplexa dynamiska miljöer och att 1-mot-1 uppgiftsdistribution ökar agenternas förmåga att genomföra experimentuppgifterna ytterligare.