Last Modified  5 October 2005

Simulations aren't evil

In the old days of `New AI', one of the things that helped define the field was the claim that simulations were evil, and that true AI could only be achieved on robots.

Why?  Two reasons:
  1. Real intelligence isn't as complex as it looks.  The apparent complexity of behavior is a reflection of the complexity of the natural environment in which the behavior is expressed.  If you take away nature and construct a simple environment, then you will either wind up putting too much complexity in your agent or you will fail to make it look very smart.
  2. In most of engineering, the hardest part of solving a problem is understanding it.  If you build a simulation, then you are claiming you understand the problem well enough to replicate that problem.  Naturally you will be able to solve the problem you've created yourself!  Simulations allow you to either deliberately or accidentally over-simplify the problems of real intelligence.


So if I know all that, why don't I think simulations are evil?  Well, basically, I've realized two more things:

  1. Nobody builds their own simulations anymore.  `Simulations' are often real problem environments that are developed by a large number of people, like the robocup simulation league or unreal tournament.  So no one can cheat; these are real problems.
  2. Robots are also simulations.  They don't have anything like the sensing or actuation of real animals, and they seldom run in realistic environments.   Arguably, a software agent that has to swim in a fluid dynamics simulation with simulated vision may be more like a real animal than a robot with wheels and a laser range finder.
I've learned a lot from working with robots (and I probably wouldn't have developed BOD without having worked on one robot platform for a year and a half).  But one of the things I learned was not to be a robot snob.

I talk about what I learned from robots more in my PhD dissertation, and why they aren't critical to AI more in a paper called "Hypothesis Testing for Complex Agents".  Both of these are available on my publications page.


page author: Joanna Bryson
back to my links page