CS 8803B - Artificial Intelligence
Lecture 1, 8/19/02
Adam Feldman
The manner in which humans interact with our environment is by recieving information via senses, then manipulating the environment through action
For example, we see a cup (sense), and pick it up (act)
Two methods of approaching AI are the "Cognitive Science" Approach and the "Intelligent Systems" Approach"
Cognitive Science: Goal is to Understand and Replicate the behavior of humans
Intelligent Systems: Goal is to just accomplish the task, without necessarily understanding how humans do it
Example - wanting to fly
Cognitive Science: Study birds and replicate their manner of flight
Intelligent System: Just build a flying machine (ie Airplane does not fly the same way as a bird)
Does not mean that certain things can not be learned from studying the bird (flight dynamics)
What is the ultimate goal of an AI Agent? Possible goal is passing the Turing Test
Turing test: Person communicates with either another person or an AI agent. If the person cannot determine that they are communicating with a machine, it passes the test
Is the Turing Test a good measure of the success of an AI Agent?
No: The test is too restrictive - it does not test all facets of intelligence
Yes: It shows that the Agent demonstrates at least some degree of intelligence
True question involves the definition of 'intelligence' as well as whether or not emulating a human actually shows intelligence
Problem with this test as an ultimate goal is that people spend too much time trying to emulate human problems (such as bad grammer/spelling, slow response times, etc)
Is the Turing Test the ultimate goal? It can be a useful measure of some things, but it is not a satisfactory ultimate goal
"Making Rational Decisions" is another possible performance measure
In this case a performance measure could be as simple as a thermostat or as complex as an android
The hope is that enough complexity will indicate AI
A Different View: What is AI?
Systems that think like humans
Cognitive Science
Systems that think rationally
Systems that act like humans
Turing Test
Systems that act rationally
Our book's approach
Rationally, as related to performance measure
Top row involves thinking
Bottom row involves acting
Left column involves emulating humans
Right column involves behaving (thinking/acting)
Why is AI Hard?
Nature of the Agent
Agent has only limited perceptual, memory, and processing capacities
No matter how much you have, it's limited
Because of these limitations, real time performance is hard
Computational Intractability
Nature of the World
World is dynamic - agents must be adaptable (change is difficult)
Not all of the world is perceptually acceptable
How do you make coffee if you can't find the coffee maker?
Data of the world is multimodal - it takes on many forms which must be able to be understood
Data can be huge - hard to make sense of it all and find the needed information
Like "looking for a needle in a haystack"
Data can be partial, relative to what is necessary to complete a goal
Nature of Computation, Knowledge, and Inferrence
Computations are local while AI problems have global constraints
Natural Language Comprehension: "Time flies like an arrow"
Looking at this sentence is a local computation (as each word is parsed), but the ambiguity of 'flies' can only be solved globally (by taking the sentence as a whole)
Vision: Seeing parts of an object is a local computation, while recognizing the whole object is a global problem
Computational logic is deductive, but many AI problems are abductive/inductive
Deductive - recognizing an effect from a cause
"If P, then Q" - given P, we know Q (P always causes Q)
Abductive - determining a cause from an effect - This is not always a sound method
"If P, then Q" - given Q, P not necessarily a valid assumption (Q could be caused by something besides P)