Abstract
M.Sc. (Computer Science)
General machine intelligence represents the principal ambition of artificial intelligence research: creating machines that readily adapt to their environment. Machine learning represents the driving force of adaptation in artificial intelligence. However, two pertinent dilemmas emerge from research into machine learning. Firstly, how do intelligent agents learn effectively in real-world environments, in which randomness, perceptual aliasing and dynamics complicate learning algorithms? Secondly, how can intelligent agents exchange knowledge and learn from one another without introducing mathematical anomalies that might impede on the effectiveness of the applied learning algorithms? In a robotic search and rescue scenario, for example, the control system of each robot must learn from its surroundings in a fast-changing and unpredictable environment while at the same time sharing its learned information with others.
In well-understood problems, an intelligent agent that is capable of solving task-specific problems will suffice. The challenge behind complex environments comes from fact that agents must solve arbitrary problems (Kaelbling et al. 1996; Ryan 2008). General problem-solving abilities are hence necessary for intelligent agents in complex environments, such as robotic applications. Although specialized machine learning techniques and cognitive hierarchical planning and learning may be a suitable solution for general problem-solving, such techniques have not been extensively explored in the context of cooperative multi-agent learning. In particular, to the knowledge of the author, no cognitive architecture has been designed which can support knowledge-sharing or self-organisation in cooperative multi-agent learning systems.
It is therefore social learning in real-world applications that forms the basis of the research presented in this dissertation. This research aims to develop a distributed cognitive architecture for cooperative multi-agent learning in complex environments. The proposed Multi-agent Learning through Distributed Adaptive Contextualization
Distributed Cognitive Architecture for Multi-agent Learning
(MALDAC) Architecture comprises a self-organising multi-agent system to address the communication constraints that the physical hardware imposes on the system.
The individual agents of the system implement their own cognitive learning architecture. The proposed Context-based Adaptive Empathy-deliberation Agent (CAEDA) Architecture investigates the applicability of emotion, ‘consciousness’, embodiment and sociability in cognitive architecture design. Cloud computing is proposed as a method of service delivery for the learning system, in which the MALDAC Architecture governs multiple CAEDA-based agents. An implementation of the proposed architecture is applied to a simulated multi-robot system to best emulate real-world complexities. Analyses indicate favourable results for the cooperative learning capabilities of the proposed MALDAC and CAEDA architectures.