Cognitive development is a complex process. Forming a new concept requires an attachment to features of the targeted object or the environment, together with an internal reasoning and mapping mechanism. Color categorization is one of the typical examples and is also a hot topic in linguistics. It involves two processes, namely color perception and color naming. Color perception is the output of the photoreceptor cells (cones) in our eyes, stimulated by the light beam reflected from an object, while color naming is based on our knowledge and language, mapping a name to the observed color.

The focus of this thesis is to study cognitive development, particularly color categorization, in a complex adaptive system which involves a population of multiple agents. The cognitive function is to be established automatically in agents via interactions without any supervision. Learning via interaction has been considered in many multi-agent systems. A typical one is the study of linguistic evolution using the naming game framework. This framework is based on interactions of pairwise agents, targeting to transfer an object’s name from one agent to another, by implementing simple update rule in agents. However, the mapping between object and name is usually ignored, while the focus is on the consensus of vocabulary, not on object classification.

In our work, a new naming game model, called Domain Learning Naming Game (DLNG), is proposed and implemented. Unlike other existing models, DLNG allows domain learning and naming, evolving in a single game. Thus, it serves as a practical model for developing cognitive functions in agents. DLNG has been applied to a network of agents for developing color categorization capability. Furthermore, the agent model has been enhanced by incorporating the concepts of subjective perception and subliminal stimulation based on human perception. Subjective perception is related to one’s viewpoint, and two colors are considered to be different if their difference is larger than an internal threshold. Subliminal stimulation refers to sensory stimulation, and colors are non-differentiable if their difference is smaller than another threshold value.

The effectiveness of DLNG has been confirmed by extensive simulations. Our results show that, with a sufficient number of interactions, consensus in color categorization can be reached in a population of agents without any supervision. A high color differentiation is also observed in the final color category. It is also interesting to point out that this category is close to the focal colors reported by major work on color languages in linguistics.

The definitions of subjective perception and subliminal stimulation allow the inclusion of uncertainty in the cognitive system of the agent. Therefore, simulations have been conducted to investigate their impacts. Results show that, if the uncertainty increases, the number of color categories decreases. Consequently, it improves the success rate of the games. However, it may not be good for the community due to the lack of vocabulary for communications. A further increase in uncertainty probably leads to the collapse of color category. On the other hand, if the uncertainty is small, meaning that agents are very precise in color categorization without much flexibility, the learning process would be slow. Degradation in game success rate is observed as it is relatively difficult to reach consensus.

Our study has also considered the inclusion of some color blind agents in the population, and investigated their potential impacts on the overall learning of the whole population. Some interesting results are observed. Finally, the impact of the underlying topology is investigated. The topology governs the possible pairwise communications between agents. Various network models, such as scale-free model with tunable clustering and WS small-world model, have been tested, and the corresponding learning processes have been compared.