Building Representations in Motivated Learning

If an intelligent system is to benefit from prior experiences, then such a system must have the ability to learn. Learning must lead to the gathering of new knowledge of increased complexity and is based on the exploration of the world and social interactions. In this chapter authors describe building representations in motivated learning, a process that is close to learning by natural systems and yields better learning results in artificial systems than reinforcement learning. An embodied agent's mission is to survive in an unfavorable environment. The agent must have needs whose fulfillment is a measure of its success – survival. Meeting these needs require physical and mental efforts, and the development of useful skills is associated with the development of intelligence. The agent's environment must provide conditions in which individuals will be subjected to pressure from an environment in which better solutions, greater skills, and broader knowledge count. The agent treats unmet needs as signals to act. The strength of these signals depends on the degree of unmet needs so that the agent can differentiate between them and compared them. Various need signals provide motivation for action and control the learning process. In complex environments, there are rules that regulate the relationships between objects. By discovering these rules, the machine gains knowledge about the environment. Knowledge is represented by building connections between neurons in semantic memory. New concepts, objects, needs, or motor skills are represented by adding new memory cells and by associating them with other concepts, actions, and needs. Whether or not a new object or idea is created in semantic memory depends on the mechanism of novelty detection. The more time an agent spends on working or playing with an object, the better it learns the object's physical properties and how to use it. The intended use of objects determines characteristic features needed to classify them. Initially, semantic memory does not store any concepts, does not know places, does not recognize any objects, and does not support any activities or motivations. New concepts or representations of objects emerge from observation and manipulation of objects. A virtual agent's semantic memory obtains symbolic representations of objects and their location or movement in the observed scene. The focus of perceptual attention may result from detection of novelty, change, movement, signal intensity, or meaning in the context of needs. Attention should be focused long enough for the working memory to evaluate how much observed object or considered plan is useful. The focus of attention must also be accompanied by the possibility of switching attention. The attention switching responds to various types of signals, from sensory stimuli through planning and monitoring of performed activities to associative activation of memory. It results from constant rivalry between these signals for attention.