May 2004

Genetic Programming (GP)

The fun thing about AI is all the biological sorts of terms used to describe cold programming algorithms, starting, of course, with the coining of the phrase "artificial intelligence" to replace the rather boring label "heuristic programming".

Genetic search algorithms follow this proud naming tradition. They are also referred to has hill-climbing search algorithms. (See Jan 2003 newsletter for discussion of genetic search and basketball schedules.) They explore a search space by mutating a given point and seeing if any of the mutations have higher values, in however height is measured, in the search space.

Genetic programming is the application of this same technique to the generation of software.

Consider that there are all sorts of formal methodologies people use for developing software. They are all good and have one thing in common. They force a discipline on the programmer, requiring him/her to think clearly and logically in the laying out of software to meet a given problem specification.

There is another approach to developing software that is much more common than one based on a formal methodology. That is to immediately start writing code and see if it works and then sort of change it when it doesn't and, without ever quite understanding exactly how it works, to keep hacking away until for some reason it finally works.

That is the development method automated by genetic programming. And why not? A lot of great programs have been written by programmers using this methodology. In fact, one definition of AI programming included the constraint that if you knew where the program was going it wasn't AI programming. A great rationalization for the pleasures of hacking.

For GP, the randomizing work follows the ideas of evolution. A number of candidate programs are randomly generated, and then each is tested against the criteria for success. The best ones are kept, the worst ones discarded. Then the best ones are mutated and/or merged with each other to create a new generation, which is then measured. The process continues until a program is generated that meets all the specifications.

Koza's Work

Dr. John Koza of Stanford is probably the best known name in genetic programming. He is now on his fourth book, Genetic Programming IV: Routine Human-Competitive Machine Intelligence (see links), which he co-authored with a number of other individuals. His earlier books lay the ground work for genetic programming, and this latest book details the current successes of the technology. I highly recommend at least reading the first chapter, which is linked to from the Genetic Programming (see links) home page.

Given that random search is the key to GP, and that the search space is large for any significant problem, one could reasonably conclude that a GP algorithm would take a long time to find a good solution. This is, in fact, the most common criticism of GP.

But machines keep getting faster. Koza and others working in the field have much faster machines today than they did when they started playing with these ideas. Using fast, parallel processing machines, they have automatically generated a number of significant applications, and by significant, they really mean significant. These are applications that in some cases replicate other patentable software; in others surpass some previous human coded algorithms; and in a few cases generated new unique patentable algorithms that should pique the interest of lawyers and entrepreneurs.

The key to success in a GP domain are to:

  • Have a good evaluation criteria that works with intermediate results, and
  • Have a good set of primitives to use to build working programs.

Here's two examples, one from RoboCup and the other my own hacking. The GP Web site has dozens more.

RoboCup

Darwin United is a team that used GP techniques to evolve a RoboCup team for the computer simulation league, which very respectively finished in the middle of the field.

The Darwin United work illustrates some of the strengths and limitations of GP. First is the need to have a reasonable function that evaluates success. Without being able to define intermediate stages, a genetic programming algorithm will thrash around forever with meaningless results.

If the criteria for success was simply to score more goals than another team, then it would be difficult to evolve better and better teams. So they had to create more detailed criteria that evaluated "good play" aside from just winning. For example, one such criteria was the ability to simply score a goal on an empty playing field.

Another problem was that it takes a GP algorithm a long time to simply generate a solution for a player to move in a straight line towards the ball. For this reason, in addition to the primitive robot moving operations that were the building blocks for the robot's behavior, higher level behaviors were also preprogrammed. These were simple behaviors the developers knew would be useful.

The Post Script paper in the links gives an excellent description of this work, although you might want to simply Google for Darwin United to get Google's automatic translation to text.

Hacking a GP Experiment

I could have continued to do more research on GP, but instead decided to start hacking away at a genetic programming experiment. It didn't get as far as I would have liked, but the experiments did illustrate some of the issues with genetic programming.

The problem I was trying to solve was the automatic generation of Prolog list utilities. These are the all too tricky bits of Prolog code that do all the useful things in the language, such as finding a member of a list or appending two lists together.

The program would start with the signature of the arguments and some test cases to specify that the program is working. For example, here were the specifications for member/2, which finds members of a list.

signature(member, [var, list]).

spec(1, member(a, [a,b,c]), true).
spec(2, member(b, [a,b,c]), true).
spec(3, member(d, [a,b,c]), fail).
spec(4, member(_, []), fail).

As with other GP applications, the first step was to define the primitives that would be used to create programs. For the case of list utilities this was constrained by the signature of the target predicate. So for member/2, the only clause heads and goals created were prolog structures named "member" with two arguments. Further, the arguments were constrained to be either lists or variables.

A Prolog predicate is made up of one or more clauses, so the mutation routines would add or delete clauses, add or delete goals for those clauses, and scramble and/or unify the arguments of the goals and heads.

After much human random hacking of my GP program, it finally generated member/2, and relatively quickly at that.

Encouraged, the next project was to generate append/3, which is a little more complex. But after a few days of frustration I finally gave up. Maybe with more patience to let the program run it would have found an answer, but my limit is letting the machine run overnight.

The problem is the evaluation function in the search space. If there are a number of small incremental degrees of success an evolving program can reach, then the GP approach is very promising, just as it would be for the human hacker, who, upon reaching some level of success can then begin to make modifications to further improve the program.

But the test cases for the list utilities contain a giant step. Some of the cases can be solved with a single clause, but the meat of the problem requires the generation of a recursive clause. There isn't some intermediate step between the single clause and the recursive clause that will help the GP towards a good solution.

This, in fact, seems to be the key human design factor in using GP to solve a problem. Instead of working on the problem itself, the developer must create an evaluation procedure that will let the GP move in the right direction. This was the critical part of Darwin United's efforts towards genetically programming a RoboCup team.

The specification used for member/2 doesn't have this smooth criteria. The program very quickly generated code that gave the right answer for test cases 1, 2, and 4:

member(A, [A|_]).

But took much longer to satisfy test case 3, member(b, [a,b,c]), because it required the addition of a recursive clause:

member(A, [_|Z]) :- member(A, Z).

The discovery of that second recursive clause is a giant leap from the simple single clause that handles most of the cases. And just as students of Prolog can spend hours thrashing about with different ways of doing the recursion, so did the program generate lots of ridiculous combinations of arguments and recursive goals in attempting to find the second clause of member/2. Such as:

member(A, [B|C]) :- member(C, [D|B]), member([B|E], C).
Without a guiding intermediate step, the genetic program can't find that second clause until it randomly generates the correct combination of goals and arguments. In the case of member/2, the second clause is simple enough that it was eventually found, but for append/3 the added complexity of the third argument just seems to have enlarged the search space to where it became very time consuming to find the right recursive clause.

[Late breaking news! The program just generated append/3 on the 261000th generation, with the 13 millionth candidate program.]

Now, with member/2, the first solutions it generated often had four or five clauses with multiple extraneous goals. In other words, the solutions were more complex than they had to be. Adding another search criteria which looked for the shortest working solution was very successful, as now the genetic algorithm had a metric it could get its teeth into, and it very quickly mutated the cumbersome first solution into the perfect elegance of the classic definition of member/2.

The code for this experiment is available for download. I think its a good start, but needs a lot of work.

FLINT - A Fuzzy Bayesian Certainty Tool Kit

LPA's (Logic Programming Associates Ltd) FLINT product includes a variety of techniques for dealing with uncertainty in a single package. It can be used to add fuzzy logic, bayesian probability, and/or certainty factors to either Prolog programs, or programs written using their Flex expert system tool.

See the links for a good description of FLINT with clear-cut examples of its use.

Inflow - AI in the Real World

Inflow is a fascinating product that maps the relationships between people in an organization. The Inflow Web site (see links) contains a number of fascinating papers describing the theory of human networks with analysis of the nature of power, economic growth, the difficulties with corporate mergers, and numerous other human network topics.

The Inflow software provides a means to measure and graph the "connectedness" of a group of people, organizations or both. All sorts of links between entities can be recorded in the package for analysis. The resulting cluster graphs provide a very intuitive feel for the dynamics of the group.

One particularly interesting paper shows the graphs for a company shortly after a merger. As you might expect, the connections between individuals of the two original companies are almost nonexistent, leading to a very clumsy and inefficient network graph.

Another paper illustrates how an analysis of a people network can lead to better management of that network. For example, to disperse information it is better to get it in the hands of the most connected individuals, rather than, say a manager of a group who might not be as connected as he/she should be.

A paper with Machiavellian overtones shows how to use a network graph to increase personal power.

Definitely a fun site to browse. And the relevance of this is the knowledge representation and reasoning engine for the Inflow software was developed using LPA's Prolog which was very well suited to this type of application.

Links

http://www.genetic-programming.org/ - The home page for genetic programming information that contains numerous references and links to other work. From the home page there are links to Koza's fourth book on GP and its first chapter, which makes very interesting reading. There are a number of other interesting links and tidbits on the site, including the two references to papers on RoboCup competitors developed using GP.

Luke, Sean. 1998. Genetic programming produced competitive soccer softbot teams for RoboCup97. In Koza, John R., Banzhaf, Wolfgang, Chellapilla, Kumar, Deb, Kalyanmoy, Dorigo, Marco, Fogel, David B., Garzon, Max H., Goldberg, David E., Iba, Hitoshi, and Riolo, Rick. (editors). Genetic Programming 1998: Proceedings of the Third Annual Conference, July 22–25, 1998, University of Wisconsin, Madison, Wisconsin. San Francisco, CA: Morgan Kaufmann. Pages 214–222. - An article describing a genetically programmed RoboCup entry that won a few games in 1997.

Andre, David and Teller, Astro. 1999. Evolving team Darwin United. In Asada, Minoru and Kitano, Hiroaki (editors). RoboCup-98: Robot Soccer World Cup II. Lecture Notes in Computer Science. Volume 1604. Berlin: Springer-Verlag. Pages 346–352. - An article describing a genetically programmed RoboCup entry that finished in the middle of the competition in 1998.

www-2.cs.cmu.edu/afs/cs/usr/astro/ public/papers/Teller_Astro.ps - A Post Script version of a paper describing Darwin United's RoboCup efforts. The paper provides clear insights into the strengths and weaknesses of GP, and a modified approach that yielded a competitive team.

http://www.idsia.ch/~juergen/gp.html - A site that claims to have the first two papers ever written on genetic programming, the second of which "re-invented" genetic programming. These pre-date Koza's "re-re-invention" of GP.

http://www.lpa.co.uk/fln.htm - An overview of LPA's FLINT, a product that supports three methods for representing uncertain information. Follow the link to more information to get some good examples of its use.

http://www.orgnet.com/ - Valdis Kreb's Web site on human networks and the Inflow product used to analyze them. The white papers are particularly good reads.

As always, feedback ideas and especially interesting links are welcome. Past issues are available at either www.ddj.com or www.ainewsletter.com.

Until next month.