Search

An important part of human intelligence, both historically and operationally, is our ability to communicate. We learn how to communicate, and maintain our communicative skills, in a society of communicators – a highly effective way to reach and maintain proficiency in this complex skill. Principles that might allow artificial agents to learn language this way are in completely known at present – the multi-dimensional nature of socio-communicative skills are beyond every machine learning framework so far proposed. Our work begins to address the challenge of proposing a way for observation-based machine learning of natural language and communication. Our framework can learn complex communicative skills with minimal up-front knowledge. The system learns by incrementally producing predictive models of causal relationships in observed data, guided by goal-inference and reasoning using forward-inverse models. We present results from two experiments where our S1 agent learns human communication by observing two humans interacting in a realtime TV-style interview, using multimodal communicative gesture and situated language to talk about recycling of various materials and objects. S1 can learn multimodal complex language and multimodal communicative acts, a vocabulary of 100 words forming natural sentences with relatively complex sentence structure, including manual deictic reference and anaphora. S1 is seeded only with high-level information about goals of the interviewer and interviewee, and a small ontology; no grammar or other information is provided to S1 a priori. The agent learns the pragmatics, semantics, and syntax of complex utterances spoken and gestures from scratch, by observing the humans compare and contrast the cost and pollution related to recycling aluminum cans, glass bottles, newspaper, plastic, and wood. After 20 hours of observation S1 can perform an unscripted TV interview with a human, in the same style, without making mistakes.

The document describes the architecture of the system being developed in the project. The principal contribution of this work is to the engineering of autonomous systems in a general fashion, i.e. independently of the target domain. This work is not “biologically inspired” and contributing to modeling natural intelligences is not an objective of the project. The main challenge set for the system is to adapt in dynamic open-ended environments with insufficient knowledge and limited resources. The system is to perform in real-time and extract knowledge from the domain it operates in. In particular, this means discovering meaningful states in the environment and learning skills by observing intentional agents in the domain. No system facing the complexity of the real world is able to learn effectively and efficiently from scratch. For each domain our system is plunged in, we hand craft a bootstrap code (called the Masterplan) that consists of the necessary initial and minimal knowledge to observe and act in the domain. The system is model-based and model-driven, meaning that its architecture is unified and consists essentially of dynamic hierarchies of models that capture knowledge in an executable form. The system is thus composed of executable models. From this perspective, learning translates into building models, integrating them into the existing hierarchies and revising them continuously. This perpetual rearranging of the internal agency of the system addresses the first objective of the HUMANOBS project:  to design an auto-reconfigurable architecture. Learning is based on model building, which in turn is driven by goals and predictions, i.e. the evaluation by the system of the observed phenomenology of the domain. In other words, the system infers what it shall do (specification) and observes ways to reach these goals (implementation). In that regard, this addresses the second objective of the project:  to build a system that can auto-generate specifications for skills and behaviors based on their observation. Learned specifications and implementations are highly context-dependent, which raises the challenge of identifying when to reuse (or refrain from reusing) learned models. Specifications and implementations are built hierarchically, which means that the system is able to reuse previously learned skills, however said skills are by design executable in parallel: this raises the challenge of coordinating (or sequentializing) the operation of the models that implement said skills. The architecture has been designed from the onset to solve these issues, and this addresses the last objective of the project:  to build behavior generation and coordination mechanisms for the reproduction and reuse of observed skills. The architecture has been designed in a principled way (each part of the architecture is based on the same architectural template), and organizes the cooperation of four continuous processes: Model Acquisition, Model Revision, Compaction (or model compression) and Reaction (reactive behavior in the domain). It is the interplay of said processes that not only ensure the viability of the whole system but also improves its performance. For example, acquiring models only requires a few examples, performs in real-time and is fast, but this process requires another process that revises models also rapidly and in real-time, and this in turn is supported by the reactive behavior of the system that pays attention only to meaningful entities and phenomena – which it has learned previously, the whole cycle having been bootstrapped by the Masterplan. This is in sharp contrast with traditional Machine Learning approaches which ignore the other cognitive processes of a system and thus, left on their own, require enormous quantities of training examples and cannot perform in real-time. A functional prototype has been developed as a proof of concept of the architecture, and the preliminary results reported in this deliverable strongly indicate that the architecture is sound and that our approach towards the engineering of autonomous systems is tractable and promising. 5/80 This prototype has been expanded, generalized and optimized furthermore into the final state of the architecture (with respect to the time frame of the project). The implemented final system satisfies the requirements of the project (although the evaluation results are presented in a separate deliverable) and, in particular, is completely domain-independent. Future developments and related research avenues have been identified and will bring the architecture beyond the requirements of this project. In the main, these developments are aimed at addressing the issues of (a) adding curiosity and imagination to the system's capabilities and, (b) controlling the autonomous bottom-up growth by means of top-down allonomic constraints.

When constructing a formal model of emotions for intelligent agents, two types of aspects have to be taken into account. First, qualitative aspects pertain to the conditions that elicit emotions. Second, quantitative aspects pertain to the actual experience and intensity of elicited emotions. In this presentation, we show how the qualitative aspects of a well-known psychological model of human emotions can be formalized in an agent specification language and how its quantitative aspects can be integrated into this model. Furthermore, we discuss several unspecified details and implicit assumptions in the psychological model that are explicated by this effort.