Jordan Pollack has worked on Artificial Intelligence using computers since 1975. In 1987, he received a Ph. D. in Computer Science from the University of Illinois. He is now a professor at Brandeis University, where he is Director of the Dynamical and Evolutionary Machine Organization, known as the DEMO Laboratory . A prolific scientist, inventor and entrepreneur, Dr. Pollack has made a few significant contributions to the fields of Artificial Intelligence and Artificial Life. Through his work on machine learning, neural networks, evolutionary computation and dynamical systems, Pollack has sought to understand the processes by which systems can self-organize and develop complex and cognitive behaviors. At DEMO, Pollack and his colleagues have applied "co-evolutionary learning" to significant problems in game playing, problem solving, search, language induction, robotics, and even educational learning across the Internet.

1.03.01 Kevin Labeau asked:
What is the time period involved in the computer selecting a tangible working model to its fruition? In other words, how long does the computer take to make a working model in situ and then become reality?

Pollack's response:
We lose patience after a couple of days! But as computers get faster, they evolve more stuff in less time, so now a run may take a couple of hours. Next year it might take minutes. The rapid prototyping machine takes almost a day to fabricate the bodies, but there are faster machines on the horizon. As an example, laser printers used to take a minute per page, and now can spit out 30 a minute.

The key thing is not the manufacturing mechanism but the constraint on what we are automatically designing so it can be built and that it can work in the real world. We expect that with the right assembly systems we can remove the human element completely for certain classes of machines.

1.03.01 Mark Dreisbach asked:
I really enjoyed the most recent program on PBS. I sat in wonderment as the computer was able to actually "learn" the best method for satisfying the question at hand. As the father of twin boys, I was wondering if any of the print outs of the Lego models would be available or a list of the particular Lego parts is listed somewhere so we might attempt to build some of your designs and put them into practice? It would be a much better use of the 1000+ Legos hanging around the house that I seem to step on in the middle of the night.

Pollack's response:
Well, the computer didn't learn the "best" way, it just gets better over time until either it stops making progress or a human gets bored waiting. We do have blueprints, but they are big things printed out on a HP plotter. You can play with Lego evolution from a web browser, on our lab's homepage, but we don't have an economical way to deliver plots to people.

1.03.01 JP asked:
I want to follow up on Alda's remark that these toy-like creatures could be the forebears of a race of super-robots which wipe out humanity. How are your robots going to follow Asimov's 3 Laws? I'm surprised PBS didn't discuss Bill Joy's warnings about self-replicating robots in the context of the show. Shouldn't we be really afraid?

Pollack's response:
If the story was about automatic design and production of fax machines, it might not seem so scary. Vending Machines, automatic teller machines and inkjet printers are robots by my definition: computer software in control of some kind of real physical machine, working 24 hours a day, moving people into other jobs. Few people want to deal with human soda vendors or to return to the days of waiting in line for 9-5 banking.

Now, if you stick your finger in a $99 inkjet printer, you will be pinched and sprayed with ink, and then it will probably break and cost more than $99 to repair. This robot has no conception of a human, little awareness of anything, except if a button is pushed, a piece of paper is loaded and where the printhead is. It has taken the industry 12 years to make printers aware of how much ink is left. It will be hundreds of years before an inkjet printer can be instructed in law.

Robots are not general purpose, but special purpose, and even though humans can build lovely machines, they end up being way too expensive for the limited work they can do. The high cost of robot design can only be recovered through mass production or in very high margin industries like software and pharmaceuticals.

Self-replication is of theoretical interest only, and has been studied for a long time, following von Neumann's work on cellular automata. We know how to do it in pure software. Until the day of artificial persons, who can work a machine shop, program a computer, order, arrange, and install the chips and motors, in their own babies, the danger is minimal. No Electro-Mechanical (EM) robot will be able to eat an old computer or fax machine to get parts for its children, because it needs highly integrated specific parts. To gain control of the means of production for an out-of-control take-over-the-world scenario a robot would need to buy General Electric, and replace all the people with computers.

I do not mean to sound facetious, because I have thought a lot about the problems. Software robots which replicate themselves, like the "love bug" Microsoft Outlook Virus, are a real threat. And nobody is considering whether all the computing power inside Cisco and Sycamore routers and switches of the Internet can mutate into an evil brain. I proposed to the government one day to set up a "Search for Extra-Terrestrial Intelligence" project, only to watch for signs of life and intelligence arising in telecommunications network dynamics.

The real issue for robotics is not an out of control self-replication problem, but whether humanity is aware of the mistakes of the past, where the power plant, the paper plant, and the automobile externalized costs and polluted the earth.

So I worry about robots which get energy from internal combustion or by eating organic lifeforms - commercial success leads to less air to breathe or food to eat. I worry about the cyborgization of animals and humans for commercial or military purposes, such that in a war, all the local animals are executed first. I worry whether robots which can perform hazardous duties enable humanity to engage in even more hazardous activities. I worry about the effect of hybrid human/robot immortality on the tenure system.

1.03.01 James Lee asked:
I have many difficulties with the theory of evolution in regards to animal and plant species. I don't understand why it is still referred to as a theory when it doesn't satisfy the basic rules of theories. Is the idea of theory being warped to fit biological evolution? I realize that your work doesn't directly approach the application of theories to animals etc., but I thought you may be able to comment on this problem.

Pollack's response:
You are right that evolution is not a theory. It is a framework, a set of principles for understanding how life emerged from non-life, and how life continues to change over time. And all evidence suggests that these are the correct, if incomplete, principles. You need a framework to ask questions scientifically, such that careful observation and experimentation lead to increased perception of the truth of the universe. If the framework does not allow questions to be asked, it is ultimately unproductive.

I think that there are many times where science reaches a dead end, and people are left with strong - almost religious - convictions but no evidence, and no means to collect it. A lot of areas of cognitive science, for example, have frameworks which initially seemed productive, but when pushed, left hard questions as magic to be answered by another field. I can think of three: AI theories which required solutions to combinatorial problems, waiting for a solution to the fundamental conundrum of computer science (does P=NP?), language acquisition theories which presume but can never know the child's biological endowment, and theories of the historical stages of consciousness which cannot be tested without a time machine.

So, while my lab's work doesn't bear on animals and plants per se, it does focus on the principles which one day may allow software to self-organize the same way as the matter, energy, and information processes we call "life". And we don't care what actually happened, or what actually exists, but only care about the process by which it came to be. Darwin is mostly right, but he did not really understand computation, dynamical systems, and game theory, because they were not developed in his day.

1.03.01 Paul Aspenson asked:
How do you explain the 2nd Law of Thermodynamics when you place this law beside your theories in evolution and natural selection?

Pollack's response:
This is a great question. In my opinion, the best definition of life is that it is a process, far from equilibrium, which wastes energy and creates structures - a local reversal of entropy. How can any system act to reverse entropy when it seems to violate the second law? The answer is that the system is open - it wasting energy given from outside. We are trying to formalize and understand this well enough that a piece of software can waste a bunch of computer cycles and create more and more of itself

We can see already how computer time is turned into knowledge, by considering a traditional game playing program with an evaluation function which estimates the goodness of any position - as it searches the tree of moves deeper and deeper, its evaluation of a board position gets better and better - in the limit, it can search all the way to the final move of a game. This basic law of AI - the knowledge-search tradeoff - is how Deep Blue beat Kasparov. Chess knowledge was formed by massive wasting of computer time using custom IBM hardware.

We often get caught up with definitions of evolution which involve sex, reproduction, survival of the fittest, animals, and so on, and obscure the underlying thermodynamic question. In our view, species, individuals, lifetimes, biological-truth-as-it-is are just arbitrary side-effects of the anti-entropic process as happened in one "run" of organic chemistry on earth. And it has run so long that it is an extreme waste of time to try to reverse engineer it. We are faced with Brains, Minds, Species, Languages, and Ecosystems, most of which are arbitrary accidents built on top of arbitrary accidents selected for a billion years.

Once such a local entropy reversal reaction is going, and there is sufficient energy to dissipate, the structures which support more and more complexification must emerge as a mathematical consequence. In other words, dissipate energy, create knowledge.

Our notion of complexity and knowledge is disputed. Kolmogorov said that the complexity of a string is just the smallest program which can generate it, leading to the conclusion that random strings are the most complex. But he didn't calculate the amount of energy dissipated in the generation. Bennet's Logical Depth and Atlan's Sophistication are more interesting measures for me because they take computer time into account.

In my lab we have started self-learning programs in game playing, optimization, language, problem-solving, design, and robotics, and each one teaches us a little more about the missing principles - which are more about the learning environment than the learning algorithm. I've moved from one field of AI to another for 25 years, and co-evolutionary learning is the first approach which I have not been able to dismiss after a few years of work. The ultimate goal is an open-ended computationally universal chain-reaction of self-organization - or at least being able to define what it is and isn't.