The Role of Non-conceptual Content in the Navigation Skills of an Artificial Agent
Abstract from thesisAgents acquire empirical knowledge from the environment through sensation. Sophisticated agents, such as humans, use concepts to understand and act within a complex natural environment. Other animals live within this same environment and are able to undertake complex behaviour, yet there is evidence that they do not have the capacity to form and hold concepts. One possible explanation is that these animals understand the world in terms of non-conceptual content. The theory of non-conceptual content can provide an interesting and useful theoretical basis for designing artificial agents that can act intelligently without the need for concepts. This theory would link the actions of artificial agents with the same explanatory mechanisms that can be attributed to animals. A computer model of an artificial world called Grid-World is used to test the searching capabilities of three different agents. The first agent searches using simple stimulus response mechanisms, a second uses a memory based on non-conceptual content and affordances and a third takes a conceptual view using a plan. The overall results from this model show that search performance improves as the representational capabilities of the agent become more sophisticated. The results also make it clear that affordances are an important element in the explanation of non-conceptual content. These are significant results, because if the behaviour of animals can be explained in terms of non-conceptual content, then there are reasons to believe that conceptual content has evolved from this simpler representational capability. If this is the case, then the evolution of conceptual content from non-conceptual content provides evidence to support a bottom up approach to artificial intelligence, and justifies the building of artificial agents using non-conceptual content, before trying to give a concept based account. Such an approach could provide the necessary simple building blocks which provide a basis for artificial agents carrying out more complex tasks, and avoid some of the 'scaling' problems currently encountered in artificial intelligence.
Return to Adrian Hopgood's page.
| Adrian Hopgood|