1,720,975 research outputs found
In the eye of the beholder: explaining behavior through mental state attribution
No full textMental-level modeling of software is a powerful abstraction in artificial intelligence that is encountered, for example, in the case of autonomous BDI-based agents. An agent of this type is equipped with goals that it attempts to achieve by selecting plans that it considers to be appropriate in light of particular beliefs that it holds. Those goals and beliefs can pertain to its 'physical' environment, but also to the mental state (goals and beliefs) of other agents. Having and maintaining a model of the mental state of others is important for believable social behavior, as is desirable in the case of human-computer interaction, but also (for example) in the case of computer-based characters that interact with each other in software applications like (serious) games. This dissertation focuses on the explanation of (partially) observed behavior by means of mental state attribution, and in doing so focuses on the perspective of a virtual beholder-entity that observes others' actions. In case those actions are of a computer-based agent, it may be so that the rules determining that agent's behavior are also available to the beholder for explanation. A scenario where this may occur is in applications that demand believability of virtual characters, and where it is feasible to give a beholder some notion of others' behavior-producing rules but not the full details about their goals and beliefs. In that case, nonmonotonic (abductive) explanatory reasoning can be employed in regard to observed actions, in order to obtain a notion of the observed agents' possible mental states. This dissertation formalizes this form of reasoning, presenting both an abductive logical account as well as a specification for its implementation in terms of answer set programming. Apart from the fact that an observed agent could have had a particular mental state, reasoning about observed actions involves the notion of 'dynamics'. In regard to the abductive account, it holds that mental states, which have been inferred as explanations for observed behavior, should be attributed to the agent in some state preceding the actions it performed. In this dissertation, propositional dynamic logic (PDL) is used as a tool for modeling those dynamics, focusing in part on the case where actions of computer-based agents are observed. Moreover, PDL is used to formalize first-order 'mindreading' - a term which is typically encountered in the literature as a theory-neutral term for referring to the explanation of behavior in mentalistic terms. In this dissertation, existing psychological models of mindreading are discussed, and employed as a basis for determining the logical format of particular patterns of mindreading. Having a formal grasp on this format can be helpful both in the elicitation of concrete instances of those patterns pertaining to behavior in particular (software) environments, as well as their implementation as an aspect of AI. This dissertation concludes with a comparison to related (logic-based) approaches to plan/intention recognition and mindreading, of which there are plenty, pointing out differences and opportunities for crossover
The logical structure of emotions
No full textEven though emotions sometimes lead us astray, there is mounting evidence from psychology and neurology that emotions have---on the whole---a positive effect on intelligent decision making and acting. Emotions help both overtly and covertly by focusing a person's attention to what is important and pruning unpromising directions of reasoning. Like humans, artificial agents---such as robots and virtual characters---have to act intelligently under resource constraints. A deep understanding of how human emotions function as innate and learned heuristics can help us in designing more effective artificial agents. Even if one does not want artificial agents to behave emotionally, it will still be useful to make these agents have knowledge of human emotions, so that they can take these into account when interacting or cooperating with humans. In order to incorporate emotions in artificial agents, a bridge must be built from psychological models of human emotions to computer science. This is done in this dissertation by capturing an emotion theory in a formal agent specification language. This formalization both serves as a foundation for the implementation of emotions in artificial agents, and enables us to formally analyze properties of the psychological model, leading to a more precise understanding of the workings of human emotion
Organizing agent organizations : syntax and operational semantics of an organization-oriented programming language
No full textThe history of software engineering in general and programming languages in particular is marked by the introduction of high-level engineering concepts, abstracting away from the rather low-level principles that are used by the machine on which the software is executed. Such high-level abstractions allow us to focus only on a few essential concepts at the same time by factoring out details. The abstractions by which we engineer complex software systems are more than often inspired by metaphorical concepts by which we understand and structure the complex world around us. A well-known example is the concept of folder for archiving our files. Introducing the concepts of agent as the metaphorical counterpart of humans and multi-agent system as the metaphorical counterpart of a society, the field of agent-oriented software engineering brings the use of abstractions in programming to an even higher level. Agents are autonomous entities that are typically programmed in terms of beliefs modeling the information they have about their world, goals denoting the situations they desire to establish and plans describing how to reach their goals. Agents participating in a multi-agent system may have been engineered by different parties with differing design objectives, implying that agents may encounter and interact with agents having conflicting goals. An illustrative example is an online marketplace on which agents interact with (unknown) parties to sell and buy their goods. Because little can be assumed about the behavior the interacting agents exhibit and nobody directly wrote the whole program encompassing the multi-agent system it is hard to predict the emerging behavior of the system as a whole. To increase the likelihood of the design objectives of the system being met, coordination media are put in place to regulate the individual agents' behavior. Some of these media are based on constructs that resemble physical every-day structures, such as a tube's entrance gate and traffic lights. Yet others use more abstract concepts we use for organizing our society, such as norms that should be followed and roles that agents can play. Getting back to the online marketplace example, agents play the role of seller and buyer, and are expected to abide by certain norms, e.g. paying the price agreed within a certain time. In this thesis we will focus on organization-oriented coordination media. In this thesis we show that research on individual agents progressed rather independently from research on agent organizations, leaving a gap between agent-oriented and organization-oriented programming. We identify what we consider the root causes underlying this gap and develop an organization-oriented programming language whose constructs accord better with the key concepts and characteristics associated with agents. Constructs for programming roles, norms and constructs for changing the norms at runtime will be investigated in particular. To understand what our programming language can (or cannot) offer, a precise description of its meaning (semantics) is indispensable. For example, to use an obligation properly, we need to know exactly when it is fulfilled or violated, and when sanctions will be imposed. Therefore, in this thesis, we formally describe the semantics of the programming constructs in a mathematically rigorous manner
Extension-based semantics of argumentation frameworks for agent interactions
No full textArgumentation plays an important part in human interaction, whether it is used for persuasion, negotiation or simply for sharing one's point of view on a certain topic. Even at the abstract level, choosing the acceptable arguments from a given set of conflicting arguments is a challenging problem, one that was given multiple solutions in the argumentation literature, in the form of argumentation semantics. An abstract argumentation framework is a directed graph that encodes the attacks between arguments. The content of the arguments is only used for deriving the attack relation and plays no further role in the abstract setting. Instead, the semantics are only defined based on the attacks between arguments. For example, one of the most intuitive constraints that can be imposed is that there should be no conflict between accepted arguments, but several other approaches are possible. On the other hand, it is difficult to say which of these constraints are the most appropriate. It is commonly agreed that there is no "best" argumentation semantics and that, instead, each of them has unique properties that make it more appropriate for some application domain or another. The first part of the thesis (Chapter 2 and Chapter 3) provides an extensive survey of existing argumentation semantics and their properties. In particular, the inclusion relations between semantics are captured in a map that can provide useful information for easily comparing novel semantics with existing ones. Such maps are augmented with argumentation frameworks that can distinguish between semantics, so as to exhibit their unique features. Argumentation frameworks and the Kripke models from modal logic are similar in the sense that they both give some meaning to a directed graph. This intuition has lead to the use of modal logic for describing argumentation semantics, first proposed by Davide Grossi in 2010. The second part of the thesis (Chapter 4 and Chapter 5) focuses on the modal definability of argumentation semantics using the global and converse modalities. We propose an algorithm that constructs models for satisfiable formulas, based on their normal form. Using this approach, we relate the formulas that describe argumentation semantics to properties that are satisfied by those semantics. This leads to a negative result showing that even a small set of properties cannot hold at the same time as modal definability. The last part of the thesis (Chapter 6) deals with the use of abstract argumentation in dynamic systems, for modeling the intentions and goals of agents, as well as the state of the environment where they are situated. The proposed model relies on the use of argumentation frameworks together with constraints on the acceptability of certain arguments. We show that the use of traditional extension-based semantics in this context is not appropriate and we propose a novel class of semantics, based on temporarily ignoring a set of arguments in order to satisfy the given constraints. Upon choosing an extension, the corresponding ignored set is to be attacked with new arguments, so as to bring about the desired extension. We use this approach to create a formal model of argumentation-based negotiation and we explain the use of the novel semantics as strategies for the agents
Explaining agent behavior in virtual training
No full textComputer games are more and more often used for training purposes. These virtual training games are exploited to train competences like leadership, negotiation or social skills. In such virtual training, a human trainee interacts with one or more virtual characters playing the trainee’s team members, colleagues or opponents. To learn from virtual training, it is important that the virtual characters display realistic human behavior. This can be achieved by human players who control the virtual game characters, or by intelligent software agents that generate the behavior of virtual characters automatically. Using intelligent agents instead of humans allows trainees to train independently of others, which gives them more training opportunities. A potential problem of using intelligent agents is that trainees do not always understand why the agents behave the way they do. For instance, virtual team members (played by intelligent agents) that do not follow the instructions of their leader (a human trainee) may have misunderstood the instructions, or disobey them on purpose. After playing the scenario, the trainee does not know whether he should communicate clearer, or give better or safer instructions. A solution is to let virtual agents explain the reasons behind their behavior. When trainees can ask their co-players to explain the motivations for their actions, they are given the opportunity to better understand played scenarios and their own performance. This thesis proposes an approach to automatically generate explanations of the behavior of virtual agents in training games. Psychological research shows that people usually explain and understand human (or human-like) behavior in terms of mental concepts like beliefs, goals and intentions. In the proposed approach, actions of virtual agents are also explained by mental concepts. To generate such explanations in an efficient way, agents are implemented in a BDI-based (Belief Desire Intention) programming language. The behavior of BDI agents is represented by beliefs, goals, plans and intentions, and their actions are determined by a reasoning process on their mental concepts. Thus, the mental concepts that are responsible for the generation of an action can be reused to explain that action. The approach can generate different types of explanations. Empirical studies with instructors, experts and novices, respectively, showed that people generally prefer explanations that contain a combination of the belief that triggered an action, and the goal that is achieved by the action. In a validation study in the domain of virtual negotiation training, subjects indicated that the agent’s explanations increased their understanding in the motivations behind its behavior. In a validation study in the domain of human-agent teamwork, subjects better understood the agent’s behavior and preferred the amount of information provided by the agent when the agent explained its behavior. Finally, the approach was extended to make agents capable of providing explanations containing predictions about the behavior of other agents. For that, the explainable agents were equipped with a theory of mind, that is, the ability to attribute mental states such as beliefs and goals to others, and based on that, make predictions about their behavior
Strategic Reasoning in Interdependence: Logical and Game-Theoretical Investigations
No full textGame theory is the branch of economics that studies interactive decision making, i.e. how entities that can reasonably be described as players of a game should behave, given their preferences and their information. Game theory is usually divided into two main branches: non-cooperative game theory, that studies the strategies that individuals should employ to reach their own goals, and cooperative game theory, that studies instead the effects of individuals joining their forces and getting the most out of their collective strategies. The present work lies somewhat in between the two sides of game theory and studies the relation between the behaviour of individuals and the behaviour of coalitions to which they belong. The first part of the thesis, called "Strategic Reasoning and Coalitional Games", studies what it means for a coalition of players to choose the best among the available alternatives, in particular what it means for a coalition to prefer a strategy above another and in what circumstances are those strategies at a coalition's disposal. Think for instance of a chess player who is setting up an attack against the opposite king. He knows that each of its pieces has invidual strengths (e.g., the knight can go to a central square, the bishop can control an important diagonal), but he is also aware that their real power lies in their combined forces (e.g., the knight and the bishop can together control a central square on an important diagonal). His reasoning starts from an individual perspective but it suddenly shifts to a coalitional one, where notions such as preferences and strategies acquire a more elaborated meaning and display specific formal properties. The thesis investigates them adopting the standard tools of logic and game-theory. The second part of the thesis, called "Strategic Reasoning and Dependence Games", elaborates further upon the study of coalitional reasoning, focusing on the network of interdependence underlying each collective decision. Consider once again the chess player who is deciding what to move. He is perfectly aware that pieces do not always perfectly and harmoniously coordinate. At times they actually obstruct each other while at other times they may even need to sacrifice themselves for their king to survive a mating attack. Their interaction displays a thick network of dependence relations (i.e. what each piece can do for the others) which strongly influences the strategies that can be played. In the classical account of cooperative game theory however this important condition is simply not taken into account. The present work bridges this gap, constructing a theory of coalitional rationality based on the resolution of its underlying dependence relations. Concretely it studies the mathematical properties characterizing those coalitions that arise from their members taking mutual advantage of each other. Finally, it relates those properties to the classical study of collective decision makin
Organizing adaptation using agents in serious games
No full textComputer games are frustrating if they are too difficult, but boring if they are too easy. This thesis proposes a framework computer games that adjust to the skill level of the user on the fly. This way games are just challenging enough for that particular person. One of the most important factors in making computer games fun is making sure that they have the right difficulty for the player. Most games can be played at different difficulties, which you can chose before you start the game. This way you need to know how good you are at the game before you start playing the game. Different people also increase their skill level at different rates. If the increase in skill level is different than the game designer expected then the game will still become too difficult or too easy for the user. Games that are able to adjust to the player skills while he is playing already exist. However these games are only able to change small elements. For example, changing the aiming accuracy of the opponents in a shooting game or giving less damage to the player if he is hurt. The framework can be used for creating games that are able to keep track of the skill level of the player and adjust accordingly. They estimate the skill level of the player by dividing the game in small parts and measuring the performance of the player on every part that is finished. Really different tasks are chosen according to the skill level of the user. For example, a game level with lots of aiming tasks if the player needs to improve its aiming skill. Game designers usually create a nice story that the player experiences while he is playing the game. But if different players get different tasks according to their skill level then the ordering of the storyline might change and possibly ruin the experience for the player. The system developed in Utrecht allow games to adjust to the user while making sure that the storyline, created by the game designer, is preserved. For example a character with a broken leg cannot suddenly walk again to make the game easier. They also makes sure that parts of the game that should be more challenging remain more challenging for the player. The game is no fun if the extra difficult opponent at the end of the game is skipped or makes it so easy that the player could easily win in the first attempt. Eventually this research will result in games that are fun to play for everyone. The games will still have the nice stories that make them exciting . But with the big difference that you can just start the game and the game will automatically make sure that the game difficulty is perfect for you
Arguing to motivate decisions
No full textDecision makers often have to make difficult trade-offs in situations where multiple aspects matter that are different by nature. For example, in a crisis scenario with a big fire in a factory, trade-offs may have to be made between the safety of the victims and personnel, and the effects on the environment. When making a complex decision, argumentation plays an important role. This thesis addresses the main research question of how a computer can use argumentation to support a person in making a complex decision. To accomplish this, the thesis focuses on answering the following two research questions: (1) how can a computer argue about why a person should prefer one decision to another, and (2) how can a computer effectively support a person in this process. Decision theory literature describes methods to decompose what a person values into general areas of concern, intermediate objectives, and specific evaluation criteria. Although expressive, the quantitative nature complicates their use in supporting complex decisions in a natural way. Argumentation literature describes methods to make decisions. Although arguing is natural, these methods cannot be used to justify preferences between decisions in more complex situations. The thesis proposes a model designed to support evaluating decisions from different perspectives, determining what perspectives should be considered, and how important these perspectives should be. This model combines decision theory with argumentation theory. In my model, value is seen from a perspective and decisions are compared from the perspectives the decision maker cares about. To understand what the decision maker cares about, his perspective is decomposed into perspectives representing the general areas of concern that he has (i.e., the values that he holds). These general areas of concern are further decomposed into intermediate objectives and specific evaluation criteria. In this way, abstract values are made concrete, which enables the computer to compare decisions on these criteria. Given an understanding of the perspectives the decision maker cares about, the computer can then use a number of argumentation schemes to justify why the decision maker should prefer one decision to another. Moreover, several argumentation schemes are proposed to justify why one argument for a decision is stronger than another. In this way, the computer can reason about what the decision maker should do. Dialogues are used to support the decision maker in a natural way. Using a dialogue, the decision maker can put forward a counterargument when he disagrees with an argument of the computer. To accomplish this, an existing dialogue system is extended such that the computer and decision maker can not only exchange arguments, but also argue about the strength of these arguments. While exchanging arguments, the computer learns more about what the decision maker values and can use this information to advance more persuasive arguments. Finally,a method is proposed that use my value model to select the most persuasive argument in the dialogue
Sensemaking software for crime analysis
No full textCriminal investigation is a difficult and laborious process that is prone to error as teams of investigators may be subject to tunnel vision, groupthink, and confirmation bias. As a result, miscarriages of justice may ensue. To overcome these problems, in the Dutch law enforcement organization, crime analysts have been given a more important role. It is now their task to critically evaluate the investigation that is going on. They have to make sense of the vast amount of evidence available in a case by generating plausible scenarios about what might have happened. Subsequently, they have to assess the quality of their scenarios and choose the best alternative. Due to the difficulty of this process, a great need exists for software that supports crime analysts in their task. However, current support tools for crime analysis do not allow analysts to record scenarios and their relation to the evidence and as a result the most important part of the analysis process remains in the analysts' minds. Therefore, they may benefit from so-called sensemaking systems that allow them to make their reasoning process explicit by visualizing scenarios and the reasons why these scenarios are supported by the evidence. Nevertheless, such sensemaking tools for crime analysis are relatively sparse and often do not incorporate a logical model of reasoning with evidence in the context of crime analysis. This thesis aims to fill this gap by proposing sensemaking software that has specifically been designed for crime analysis. Such a tool should be rationally well-founded, natural, useful, usable, and effective. To this aid, a proof-of-concept application called AVERs (Argument Visualization for Evidential Reasoning based on stories) was built that implements a rationally well-founded and natural model of the reasoning that takes place in crime analysis. In this way a standard of rational reasoning is encouraged and errors may be reduced. Using AVERs analysts are able to create visual representations of scenarios and evidential arguments. Scenarios are represented as causal networks of events, while evidential arguments are arguments based on the evidential data in the case. Such arguments are based on argumentation schemes that often come with critical questions. These questions make the analysts more aware of possible sources of doubt and encourage them to critically examine the evidence. Evidential arguments can be used to support or attack scenarios with the available evidence. In this way, this software allows the analysts to reason about scenarios and to critically evaluate them. Moreover, it provides features that can be used to compare alternative scenarios. A series of empirical studies has confirmed that the design and implementation of AVERs fulfills all five criteria to a certain degree. This means that it is useful to crime analysts and satisfies their desires, while it may improve their analysis of the case and the communication of their results to the investigators working on the case, and ensures that rational analyses are performed. Therefore, through this software in the future biases in the crime analysis process may be avoided
Insights in reinforcement rearning : formal analysis and empirical evaluation of temporal-difference learning algorithms
No full textA key aspect of artificial intelligence is the ability to learn from experience. If examples of correct solutions exist, supervised learning techniques can be used to predict what the correct solution will be for future observations. However, often such examples are not readily available. The field of reinforcement learning investigates methods that can learn from experience when no examples of correct behavior are given, but a reinforcement signal is supplied to the learning entity. Many problems fit this problem description. In games, the reinforcement signal might be whether or not the game was won. In economic settings, the reinforcement can represent the profit or loss that is eventually made. Furthermore, in robotics it is often easier to specify how well the robot is doing than it is to find examples of good behavior beforehand. An advantage of reinforcement learning is that the designer of the system need not know what good solutions to a problem may be. Rather, the system will find good solutions by trial and error. Of particular interest to us are model-free temporal-difference algorithms. These algorithms do not use experiences to build an explicit model of the environment, but construct an approximation of the expected value for each possible action. The values can then be used to construct solutions. These methods are computationally efficient, easy to implement and often find solutions quickly. Additionally, in many settings it is easier to find a good policy to select actions than to model the whole environment and then to use this model to try to determine what to do. In this dissertation, we analyze several existing model-free temporal-difference algorithms. We discuss some problems with these approaches, such as a potentially huge overestimation of the action values by the popular Q-learning algorithm. We discuss ways to prevent these issues and propose a number of new algorithms. We analyze the new algorithms and compare their performance on a number of tasks. We conclude that it depends highly on the characteristics of the problem which algorithm performs best. We give some indications on which algorithms are to be preferred in different problem settings. To solve problems with unknown characteristics, we propose using ensemble methods that combine action-selection policies of a number of different entities. We discuss several approaches to combine these policies and demonstrate empirically that good solutions can reliably be found. Additionally, we extend the idea of model-free temporal-difference algorithms to problems with continuous action spaces. In such problems, conventional approaches are not applicable, because they can not handle the infinite number of possible actions. We propose a new algorithm that is explicitly designed for continuous spaces and show that it compares favorably to the current state of the art
- …
