Our prototypical AI system in this class is a single intelligent agent. The agent acts alone. The agent has a knowledge base (facts about the world) and some algorithms to operate on that knowledge base (searching algorithms, for example).
This “intelligent agent” metaphor becomes stretched when we look at certain other forms of intelligent systems.
Some systems utilize a collection of communicating agents. Each agent typically has a specialization: some task that it performs better than other tasks. A group of communicating agents come together to solve complex problems, with each agent solving just a small part of the problem.
The “intelligent agent” metaphor is stretched here because no single agent achieves the system’s goal; no single agent possesses enough intelligence to do so. Designing a multi-agent system requires thinking both about each individual agent and how they come together.
Marvin Minsky, a founder of the field of AI, believes that the mind takes this form. He calls his model the “society of mind.” In this model, the mind consists of many independent agents that each perform one small task (such as recognizing shapes, or responding to hot things, or retrieving a memory, or coordinating among other agents, etc.). Minsky was trying to develop a theory of natural intelligence, so an important feature of the model is that no single agent has a mind; each agent is too simplistic. Rather, the idea of the “mind” comes from the quick, widespread, and possibly chaotic interaction of the agents.
Minsky’s “society of mind” model is an example of emergent intelligence. An even simpler, and demonstrably effective, form of emergent intelligence is ant and honey bee behavior. Each ant or bee is very simple. Each effectively performs a random walk until it comes upon some morsel of food. The ants do not even communicate the location of the discovery; rather, other ants (randomly) find themselves on the same path as a prior ant (paths are indicated by pheromones) and follow that path rather than continue their random walks. This second ant releases more pheromones (as they all do), which reinforces the path and increases the probability that another ant will follow it should another ant randomly discover it. Honey bees, on the other hand, communicate a small amount of information about their finding by performing the “waggle dance.” This dance communicates the quality of a flower and its direction and distance.
Studied at an individual level, ants and honey bees are too simple to be effective models for AI systems. They have virtually no memory and little variation in their behaviors. However, when hundreds or thousands of ants or honey bees are brought together, an intelligent system emerges.
We’ll practice with ant intelligence and look at honey bee intelligence. Some modern AI applications that employ ant/bee intelligence include network routing, airplane routing, computer vision and pattern recognition, job scheduling, etc.
Traditional AI systems are cathedral-like constructions, with lots of interconnected subsystems, complex rules for their interaction, libraries of knowledge and formal procedures for using such knowledge. This makes them somewhat brittle, since the interactions among subsystems are so sophisticated.
However, Rodney Brooks has shown that an agent need not be intelligent in order to perform well in complex environments. Must like insects, his robots do not keep a library of knowledge and perform no search algorithms. They just continually respond to data obtained from their sensors. In a sense, all their knowledge about the world is kept in the world itself, not kept in the robot’s mind. If the robot is designed to navigate towards the light, it does not try to keep an internal map of where it is and where the light is; rather, it just continually moves an inch towards whatever direction it currently senses the light.
Brooks was a founder of iRobot, the company that created the Roomba and builds many military robots. Most of their robots do much more than seek light. By employing many fast, accurate sensors and a fast computation cycle, their (huge) robots can travel very quickly over very rough terrain while following a group of humans who direct the robot with radio signals.