An Emotion-Based “Conscious” Software Agent Architecture

Abstract

Evidence of the role of emotions in the action selection processes of environmentally situated agents continues to mount. This is no less true for autonomous software agents. Here we are concerned with such software agents that model a psychological theory of consciousness, global workspace theory. We briefly describe the architecture of two such agents, CMattie and IDA, and the role emotions play in each. Both agents communicate with humans in natural language, the first about seminars and the like, the second about job possibilities. IDA must also deliberate on various scenarios and negotiate with humans. In CMattie emotions occur in response to incoming stimuli from the environment and affect behavior indirectly by strengthening or weakening drives. In IDA the emotions are integrated with the “consciousness” mechanism, and bidirectionally connected with all the major parts of the architecture. Thus, emotions will affect, and be affected by, essentially all of the agent’s disparate cognitive processes. They will thus play a role in essentially all cognitive activity including perception, memory, “consciousness”, action selection, learning, and metacognition. These emotional connections will provide a common currency among the several modules of the agent architecture. These connections will also allow for the learning of complex emotions. The emotions serve to tell the agent how well it’s doing.

Supported in part by ONR grant N00014-98-1-0332