Towards Representational Autonomy of Agents in Artificial Environments

Autonomy is a crucial property of an artificial agent. The type of representational structures and the role they play in the preservation of an agent’s autonomy are pointed out. A framework of self-organised Peircean semiotic processes is introduced and it is then used to demonstrate the emergence of grounded representational structures in agents interacting with their environment. 1 Autonomy and Representations in Strong Agency There is an interesting interdependence between the three fundamental properties of interactivity, intentionality and autonomy which are used to describe an agent. As it is suggested in [1], there is no function without autonomy, no intentionality without function and no meaning without intentionality. The circle closes by considering meaning (representational content) as a prerequisite for the maintenance of system’s autonomy during its interaction. Moreover, the notion of representation is central to almost all theories of cognition, therefore being directly and indirectly connected with fundamental problems in the design of artificial cognitive agents [2], at the pure cognitivistic framework as much as at the embodied and dynamic approaches [3]. Although an embodied agent seems to be able to handle very simple tasks with only primitive stimulus-response actions, its cognitive capabilities cannot scale to tackle more complex phenomena. These and other problems are evidences that the use of representations, even in reflexive behaviors, becomes essential [4]. However, representations should not be generic, context-free and predetermined, but they should be an emergent product of the interaction between an agent and its environment [2]. 2 Emergent Representations via Self-organised Semiotic Processes Self-organised and embodied systems admit no functional usefulness to representations. Based on the abovementioned, the incorporation of a process to support the vehicle of the representation which carries internal information about an external state seems imperative. This process should give the interactive dimension to the self-organising system and furthermore, it should correspond to the embedded structure of emergent representations. Peircean semiosis [5] can be seen as the process which drives the system into meaningful interaction. In the proposed framework, intelligence is not considered as an extra module, but as an asset emerging from the agent’s functionality for interaction and the aim is the unification of the modality of interaction, perception and action with the smallest possible number of representational primitives. The present attempt is in correspondence with contemporary works in AI, such as [6] and [7]. In the present paper, there is an attempt to design a more generic architecture which will integrate aspects of self-organisation and embodiment with Peircean semiotics. There is in no way a demonstration of a totally autonomous system, but the introduced architecture overcomes the symbol-grounding Towards Representational Autonomy of Agents in Artificial Environments 2 problem, which is the fundamental obstacle for the frame problem, and by doing so, it introduces a type of representational structures that are integrated into the functional structure of the artificial agent. 2.1 The Structure of Peircean Signs The basic structural element of the proposed framework is the semiotic component. A possible representation is to use a frame-like structure, and to let individual slots express the respective qualities (qualisigns) of the object they represent. For indexing and interpretation purposes, two more slots should be reserved to describe the unique id of the component and the type of data it holds. In the case of artificial environments, possible objects that can be represented in the agent’s knowledge base using semiotic components are entities: the individual visual elements that exist as geometries in the environment. The semiotic component should possibly contain their spatial properties (e.g. translation, rotation, bounding box size) and other custom qualities that better describe their nature. Semiotic components could also describe: relations, i.e. spatial (e.g. near), structural (e.g. part-of) or other relations between entities, situations, i.e. a collection of objects and relations between them that describes (part of) the environment and actions, i.e. preconditions (described as the initial situation), performance (series of motor commands) and effects (changes between initial and final situation). The slots can contain either crisp values or sets. In the latter case, the component describes not just one object but a category (legisign). 2.2 Self-organised Peircean Semiotic Processes The abstract architecture rising from the interaction of a self-organised system with its environment based on Peircean semiotic processes is shown in Fig. 1. A detailed analysis of the architecture is given in [8]. Fig. 1. An agent engaging in self-organised semiotic processes with the environment. As a first step towards a computational methodology for implementing the proposed framework, an example has been set up, where agents are wandering around an environment and try to learn simple actions. Each agent has its own abilities concerning perception and action and initially it has no representational structures regarding possible actions. A perception mechanism, which is constantly being informed by the environment, creates the Immediate Objects (IO) as components that will drive the semiotic process. These are stored in the short term memory, which agents are constantly examining and comparing to their representational structure to try and detect any surprising phenomena, i.e. objects that they cannot categorize. In the implemented example, the semiotic components describe entities, IO II Self-anisation of FI DO DO DO DI Cognitive System Cognitive System Cognitive System DO ENVIRONMENT Abduction Deduction Induction