An Abstraction Architecture for Cognitive Agents
The project"s vision is to develop an architecture for cognitive agents and to validate it in a robotic embodiment in an unknown outdoors environment. The architecture will include major building blocks found in human cognitive processes and it will integrate the cycle of perception-knowledge acquisition-abstraction-reasoning-action generation. Knowledge acquisition will be supported by a suitable concept system with a corresponding abstraction mechanism. The concept system is the representation of the knowledge that the agent possesses of its environment and itself. Objects, relations, goals, context information, and solution strategies are considered as knowledge about a situation. The abstraction mechanism is responsible for creating and organising a hierarchy of concepts while the reasoning process operates on the concept system in order to make inferences for virtual actions and select the one that will realise the greatest reward. The architecture will include attention control as a means of handling complexity, prioritising responses, detecting novelty and creating new goals. Both sensory and motor attention will be used. A goals-oriented computational model will allow the fusion together of user tasks with tasks originating from the agent. A goals generation system will enable the agent to produce its own goals. Reinforcement learning will provide the means by which the agent learns solution strategies for the satisfaction of a goal. The loop closes by having new actions modifying the current knowledge through perception. The architecture will be implemented in a robotics application, namely of robot navigation in unknown outdoors environment but it will be in no way specific to this domain. Different outdoors environments will be used for testing.