The Future of Artificial Intelligence John Paxton Montana State University August 14, 2003.

The Future of Artificial Intelligence

John Paxton

Montana State University

August 14, 2003

Bannack

What makes AI difficult?

• Different problems have inherently different complexities to solve.

The Sorting Problem

• Input: 2 4 6 7 5 3 1

• Output: 1 2 3 4 5 6 7

Selection Sort

• Step 1: 2 4 6 7 5 3 1

• Step 2: 2 4 6 1 5 3 7

• Step 3: 2 4 3 1 5 6 7

• Step 4: 2 4 3 1 5 6 7

• Step 5: 2 1 3 4 5 6 7

• Step 6: 2 1 3 4 5 6 7

• Step 7: 1 2 3 4 5 6 7

Selection Sort

• If there are n items to sort, selection sort takes O(n2) time

• What does this mean? If we double the size of the input, we can expect the algorithm to take four times as long.

Quicksort

• O(n log2 n)

2 4 6 7 5 3 1

1 4 6 7 5 3

3 6 7 5

5 7

Quicksort

n n log2 n n2

10 33.22 100

100 66.44 10000

1000 99.66 1,000,000

10000 132.88 100,000,000

Sorting

• It can be proven that sorting n numbers based on comparisons has a best case of O(n log n).

• Thus, the inherent complexity of sorting is O(n log n), even though worse algorithms such as selection sort exist.

The Class P

• P = Polynomial

• Any problem whose inherent complexity is O(np) where p is a constant is in the class P.

• Problems that are in P typically are practical to solve on computers.

Travelling Salesperson Problem

• Starting in City A, what is the shortest circuit that visits cities B, C, and D?

• A – B – C – D – A• A – B – D – C – A• A – C – B – D – A• A – C – D – B – A• A – D – B – C – A• A – D – C – B - A

TSP

• In the preceding problem, there were 4 cities and 3! possible solutions

• In general, if there are n cities, one must consider (n-1)! possibilities.

• (n-1)! is not O(np) for any fixed p. (n-1)! is in the EXP class.

• Each problem in the EXP class is O(pn) for some fixed p.

Comparison

n n2 (n-1)!

5 25 24

10 100 362,880

15 225 8.7E10

20 400 1.2E17

The Class EXP

• As you can see from the preceding table, problems that are in the class EXP do not have practical solutions on computers

Relevance to AI

• Unfortunately, many interesting problems in AI are in the class EXP.

• For example, the TSP problem.

Satisficing

• What can be done?

• Instead of settling for the optimal answer, look for a “pretty good” solution instead. This technique is also known as satisficing.

Satisficing Example

Heuristics

• A “heuristic” is a rule-of-thumb that works in practice, but has no guarantee of being optimal.

Water Jug Problem

• Place 6 liters of water in the 8 gallon jug in as few steps as possible

8 3

Water Jug Problem

• Place 8 liters of water in the 10 gallon jug in as few steps as possible

10 4

Water Jug Problem

• Place 10 liters of water in the 15 gallon jug in as few steps as possible

15 5

Past AI Predictions

• Game Playing. Researchers thought that AI chess playing programs would beat the best humans by 1970.

• Machine Translation.– The spirit is willing, but the flesh is weak.– The whisky is strong, but the meat is rotten.

Objections to AI

• Theology• Heads-in-the-Sand• Mathematical• Self Awareness• Capability X is lacking (e.g. enjoy ice cream)• Lady Lovelace’s objection• Continuity of nervous system• Informality of behavior (no rules)• ESP

The Future

• Consumer Robots

The Future

• Gastrobots (University of South Florida)

• Sustain themselves by eating naturally occurring foods

The Future

• COG, a robot at MIT

• Track eye movement• Recognize faces• Grab objects• Hear a rhythm, play it

back on drums

The Future

• Art – Raymond Kurzweil’s screensaver program, Aaron

• Poetry

• Music

The Future

• Natural Language

• Charles Schwab incorporates iPhrase at its web site to allow users to use natural language to ask questions. For example, “Which of these stocks has the highest revenues?”

The Future

• Products that do one thing well.

• For example, Continental Divide Robotics has developed a system based on GPS that can locate any person or any object anywhere in the world and notify a user if it is “out of bounds”. This could help a parent monitor a child, for example.

The Future

• Companionship

• At Microsoft, a product is under development that learns about you. Who is important to you? Are you busy? The product can then monitor incoming e-mails and phone calls.

The Future• Virtual Reality

• Haptek, People Putty

• Create your own 3-D interactive characters

The Future

• Computers will get faster

• Software will get better

• AI will creep closer to human capabilities (search, learning, knowledge representation)

The Future

• There are lots of potential benefits!

• There are certainly some potential drawbacks!

• Most AI researchers believe humans will stay in control

Questions?