Embodied Cognitive Science - General Principles of Intelligent Behavior

General Principles of Intelligent Behavior

In the formation of general principles of intelligent behavior, Pfeifer intended to be contrary to older principles given in Traditional Artificial Intelligence. The most dramatic difference is that the principles are applicable only to situated robotic agents in the real world, a domain where Traditional Artificial Intelligence showed the least promise.

Principle of Cheap Design and Redundancy: Pfeifer realized that implicit assumptions made by engineers often substantially influence a control architecture's complexity. This insight is reflected in discussions of the scalability problem in robotics. The internal processing needed for some bad architectures can grow out of proportion to new tasks needed of an agent.

One of the primary reasons for scalability problems is that the amount of programming and knowledge engineering that the robot designers have to perform grows very rapidly with the complexity of the robot's tasks. There is mounting evidence that pre-programming cannot be the solution to the scalability problem ... The problem is that programmers introduce too many hidden assumptions in the robot's code.

The proposed solutions are to have the agent exploit the inherent physics of its environment, to exploit the constraints of its niche, and to have agent morphology based on parsimony and the principle of Redundancy. Redundancy reflects the desire for the error-correction of signals afforded by duplicating like channels. Additionally, it reflects the desire to exploit the associations between sensory modalities. (See redundant modalities). In terms of design, this implies that redundancy should be introduced with respect not only to one sensory modality but to several. It has been suggested that the fusion and transfer of knowledge between modalities can be the basis of reducing the size of the sense data taken from the real world. This again addresses the scalability problem.

Principle of Parallel, Loosely-coupled Processes: An alternative to hierarchical methods of knowledge and action selection. This design principle differs most importantly from the Sense-Think-Act cycle of traditional AI. Since it does not involve this famous cycle, it is not affected by the Frame problem.

Principle of Sensory-Motor Coordination: Ideally, internal mechanisms in an agent should give rise to things like memory and choice-making in an emergent fashion, rather than being prescriptively programmed from the beginning. These kinds of things are allowed to emerge as the agent interacts with the environment. The motto is, build fewer assumptions into the agent's controller now, so that learning can be more robust and idiosyncratic in the future.

Principle of Ecological Balance: This is more a theory than a principle, but its implications are widespread. Its claim is that the internal processing of an agent cannot be made more complex unless there is a corresponding increase in complexity of the motors, limbs, and sensors of the agent. In other words, the extra complexity added to the brain of a simple robot will not create any discernible change in its behavior. The robot's morphology must already contain the complexity in itself to allow enough "breathing room" for more internal processing to develop.

The Value Principle: This was the architecture developed in the Darwin III robot of Gerald Edelman. It relies heavily on connectionism.