Artificial Intelligence (AI)
It's a lot of different things to a lot of different people. Let's go through a few things that AI is thought to be and situate them within the broader picture of AI.
One thing it could be is "Making computational models of human behavior". Since we believe that humans are intelligent, therefore models of intelligent behavior must be AI. There's a great paper by Turing who really set up this idea of AI as making models of human behavior (link). In this way of thinking of AI, how would you proceed as an AI scientist? One way, which would be a kind of cognitive science, is to do experiments on humans, see how they behave in certain situations and see if you could make computers behave in that same way. Imagine that you wanted to make a program that played poker. Instead of making the best possible poker-playing program, you would make one that played poker like people do.
This is a stronger and more constrained view of what the enterprise is. It is not enough to make a program that seems to behave the way humans do; you want to make a program that does it the way humans do it. A lot of people have worked on this in cognitive science and in an area called cognitive neuroscience. The research strategy is to affiliate with someone who does experiments that reveal something about what goes on inside people's heads and then build computational models that mirror those kind of processes.
A crucial question is to decide at what level to mirror what goes on inside people's heads. Someone might try to model it a very high-level, for example, dividing processing into high-level vision, memory, and cognition modules; they try to get the modularity to be accurate but they don't worry too much about the details of how the modules are implemented. Other people might pick the neuron as a kind of computational unit that feels like it's justified in terms of neurophysiology, and then take that abstract neuron and make computational mechanisms out of it. It seems justified because we know that brains are made out of neurons. But then, if you talk to people that study neurons, you find that they argue a lot about what neurons can and can't do computationally and whether they are a good abstraction so maybe you might want to make your models at a lower level. So, it’s hard to know how to match up what we know about brains with computational models.
Software that gathers information about an environment and takes actions based on that information
- a robot
- a web shopping program
- a factory
- a traffic control system…
The Agent and the Environment
So, how do we think about agents? How can we begin to formalize the problem of building an agent? Well, the first thing that we're going to do, which some people object to fairly violently, is to make a dichotomy between an agent and its environment. There are people in AI that want to argue that this is exactly the wrong thing to do; that I shouldn't try to give an account of how an agent works by separating it from the world it works in, because the interface is so big and so complicated. And that may be right. That I can't get exactly right a description of how the agent needs to operate in the world by separating it from the world. But, it gives me a kind of leverage in designing the system that I need right now because I, as the designer of the system, am not smart enough to consider the system as a whole.