The exploration of unknown environments by affective agents

Abstract

In this thesis, we study the problem of the exploration of unknown environments populated with entities by affective autonomous agents. The goal of these agents is twofold: (i) the acquisition of maps of the environment – metric maps – to be stored in memory, where the cells occupied by the entities that populate that environment are represented; (ii) the construction of models of those entities. We examine this problem through simulations because of the various advantages this approach offers, mainly efficiency, more control, and easy focus of the research. Furthermore, the simulation approach can be used because the simplifications that we made do not influence the value of the results. With this end, we have developed a framework to build multi-agent systems comprising affective agents and then, based on this platform, we developed an application for the exploration of unknown environments. This application is a simulated multi-agent environment in which, in addition to inanimate agents (objects), there are agents interacting in a simple way, whose goal is to explore the environment. By relying on an affective component plus ideas from the Belief-Desire-Intention model, our approach to building artificial agents is that of assigning agents mentalistic qualities such as feelings, basic desires, memory/beliefs, desires/goals, and intentions. The inclusion of affect in the agent architecture is supported by the psychological and neuroscience research over the past decades which suggests that emotions and, in general, motivations play a critical role in decision-making, action, and reasoning, by influencing a variety of cognitive processes (e.g., attention, perception, planning, etc.). Reflecting the primacy of those mentalistic qualities, the architecture of an agent includes the following modules: sensors, memory/beliefs (for entities - which comprises both analogical and propositional knowledge representations -, plans, and maps of the environment), desires/goals, intentions, basic desires (basic motivations/motives), feelings, and reasoning. The key components that determine the exhibition of the exploratory behaviour in an agent are the kind of basic desires, feelings, goals and plans with which the agent is equipped. Based on solid, psychological experimental evidence, an agent is equipped in advance with the basic desires for minimal hunger, maximal information gain (maximal reduction of curiosity), and maximal surprise, as well as with the correspondent feelings of hunger, curiosity and surprise. Each one of those basic desires drives the agent to reduce or to maximize a particular feeling. The desire for minimal hunger, maximal information gain and maximal surprise directs the agent, respectively, to reduce the feeling of hunger, to reduce the feeling of curiosity (by maximizing information gain) and to maximize the feeling of surprise. The desire to reduce curiosity does not mean that the agent dislike curiosity. Instead, it means the agent desires selecting actions whose execution maximizes the reduction of curiosity, i.e., actions that are preceded by maximal levels of curiosity and followed by minimal levels of curiosity, which corresponds to maximize information gain. The intensity of these feelings is, therefore, important to compute the degree of satisfaction of the basic desires. For the basic desires of minimal hunger and maximal surprise it is given by the expected intensities of the feelings of hunger and surprise, respectively, after performing an action, while for the desire of maximal information gain it is given by the intensity of the feeling of curiosity before performing the action (this is the expected information gain). The memory of an agent is setup with goals and decision-theoretic, hierarchical task-network plans for visiting entities that populate the environment, regions of the environment, and for going to places where the agent can recharge its battery. New goals are generated for each unvisited entity of the environment, for each place in the frontier of the explored area, and for recharging battery, by adapting past goals and plans to the current world state computed based on sensorial information and on the generation of expectations and assumptions for the gaps in the environment information provided by the sensors. These new goals and respective plans are then ranked according to their Expected Utility which reflects the positive and negative relevance for the basic desires of their accomplishment. The first one, i.e., the one with highest Expected Utility is taken as an intention. Besides evaluating the computational model of surprise, we experimentally investigated through simulations the following issues: the role of the exploration strategy (role of surprise, curiosity, and hunger), environment complexity, and amplitude of the visual field on the performance of the exploration of environments populated with entities; the role of the size or, to some extent, of the diversity of the memory of entities, and environment complexity on map-building by exploitation. The main results show that: the computational model of surprise is a satisfactory model of human surprise; the exploration of unknown environments populated with entities can be robustly and efficiently performed by affective agents (the strategies that rely on hunger combined or not with curiosity or surprise outperform significantly the others, being strong contenders to the classical strategy based on entropy and cost).