Dr. Andrew L. Nelson - “Artificial Evolution and Unnatural Selection”
February 23rd, 2008 by Richard Leis, Jr.[Back to Meetings]
http://hplusclub.com/tucson/meeting20080307
Dr. Andrew L. Nelson - "Artificial Evolution and Unnatural Selection"
Date and Time: Friday, March 07, 2008 at 4:00pm until 6:00pm MST
Location: Ventana Room, 4th Floor, Student Union Memorial Center, University of Arizona - 1303 E. University Blvd., Tucson, AZ 85721-0017, USA | Google Maps
Description:
Speaker: Dr. Andrew L. Nelson was born Laramie Wyoming in 1967. He received his B.S. degree with concentration in Computer Science from the Evergreen State College in Olympia Washington in 1990. He received his M.S. in Electrical Engineering from North Carolina State University in 2000. He received his Ph.D. in Electrical Engineering at the Center for Robotics and Intelligent Machines (CRIM) at North Carolina State University in 2003. Between 2003 and 2005 he was a visiting researcher at the University of South Florida. Currently he is a researcher at Androtics LLC, Tucson AZ and Santa Cruz CA. His main interests are in the fields of fully autonomous robot control, bio-inspired robot control and evolutionary robotics. His robotics work has included applying artificial evolution to synthesize controllers for swarms of autonomous robots as well as the development of a fuzzy-logic based controllers for robot navigation. He pursues work in artificial neural networks, genetic algorithms and soft computing related to autonomous machine control. He has also conducted research in diverse fields including electric machine design and molecular biology.
Abstract: In recent years researchers interested in creating artificial life forms have turned their attention toward techniques involving artificial evolution. Evolutionary computing (EC) is a large, and rapidly growing area of research focused on exploiting computational processes that mimic natural evolution. Evolutionary computing techniques are applied to a wide range of optimization, classification and control problems. Although the particulars of implementation vary greatly, most evolutionary computing applications employ some variation on the following steps: A population of potential solutions to a particular problem (often referred to as candidate solutions) is generated. The individuals in this population are tested to determine how well each of them solves the problem, and the better performing solutions are selected. These better performing solutions are then altered by some stochastic process and then returned to the larger population of solutions to replace the most poorly performing individuals. The sequence of testing, selection, alteration and replacement is then repeated until a suitably proficient solution arises.A key component of this process - one might argue, the key component - is the measurement of fitness of the evolving candidate solutions. For many evolutionary computing applications, there are well-defined and efficient methods for determining the fitness of a given solution. However, determining fitness by using a function, or any type of measurement is unnatural.
Nature applies no particular criteria for survival. At a high level, the phrase "survival of the fittest" seems to confer some meaning to the concept of selective pressure in natural evolution, but in fact, this phrase reduces to little more than a truism: "survival of those that survive". The universe takes no action at any level beyond simple iteration of fundamental physical law. The persistence of structures or patterns of matter, including rocks, stellar material, and life-forms boils down to simple possibility: A pattern that is possible, given physical law and past configurations of matter, will exist, regulated only by stochastic factors embodied in the fundamental fabric of the universe. Things do not exist because they are better at existing. They exist because they are possible.
Fitness functions or objective functions work well for optimization and generation of particular well-defined solutions to many traditional problems, as has been demonstrated by the success of evolutionary computing. But what is the effect of attempting to use explicit fitness functions when trying to evolve life-like entities?
Presentations
- [PowerPoint] Artificial Evolution and Unnatural Selection by Dr. Andrew L. Nelson
- [AVI Movie] run3_3_21_03_movie.avi
- [AVI Movie] pop13_giant_win2_movie.avi
Meeting Notes
Androtics, LLC Machine Learning Presentation h+
1. Case Study: Evolving Robot Brains
2. Problems with Artificial Evolution
3. Objective function-free systems
Evolutionary Robotics (ER)
- Synthesis of controllers to perform a certain task.
- Robots in mention
- EvBots
- Tactile sensors
- Cameras
- Colored shells
- Building basic logic to do tasks
- Move around in a reconfigurable maze, walls are higher than robots to occlude vision
- Controlled by "monolithic neural networks"
- 60 neurons with connections to other neurons
- EvBots
- Control Scheme
- Image -> Color identification processing -> real valued numerical arrays -> artificial neural net 150 inputs 2 outputs
- Intuition leaked into the system somewhere along the lines, isn't entirely machine derived
- Different from simply optimizing a system
- Population initialization P(K=0)
- performance of controllers instantiated in robots
- fitness evaluation of each p in P based on performance in environment
- re-order P based on fitness data
- propagate unaltered and altered copies back into the population P(K+1)
- Robotic Capture the Flag (CTF)
- one team of robots versus another
- attempt to find a goal object (flag)
- 20 networks playing 1 game against another network
- walls aren't in the same place each time
- goals are in a random place
- Fitness(#WonGames)
- Solutions always compete because they are ranked
- Fitness landscape changes each time (because the fitness function changes each scenario run, so there isn't some very low goal)
- Modified right-hand mouse search strategy
- Achieved a level of performance after a certain time
- 240 games and simulations were generally run
- The simulation is ended as soon as one of the robots reaches the "goal" objective.
- Used a physical system as well as software emulated system to run the simulation
- With partially processed sensors you need to simulate parts of the environment, but not what it would look like
- So some simulations are run through hardware-emulation and some are run in a physical environment.
- Physical simulation and Software simulation are both appealing visually
- Some bots seem to get stuck.
- Move in steps (fixed discrete time intervals)
The Problem with Artificial Evolution
- What is the force that drives evolution in nature
- Survival of the fittest
- What is the force that drives evolution in nature
- Simple physical law drives the formation and development of life
- Survival of the fittest is a truism (the fittest survive, an there isn't much there, the argument is circular)
- Doesn't have the power it should have in terms of description
- Life innovates: Encoding of complex patterns into data structure
- Self replication
- Makes the nearly impossible, possible
- What organisms do is impossible (almost) for the rest of the lifeless universe, but organisms still reduce to physical law entirely
- Is life a self-replicating system
- What is the definition of life?!
- I don't know what life is, but I know it when I see it
- The force against entropy is life? Life climbs a different path in entropy
- Somehow or other it oils down to physics, nothing is stamping out the bad ones and letting good ones live
- Crystals?
- Fire?
-
- Predators versus Prey (if you consider earth as being 1 system, then these are all part of that system)
- We see survival of the fittest, yet it isn't really there
- So... what is the problem?
- The problem is what happens when we try to build artificial evolutionary systems
- We try to artificially install a survival of the fittest mechanism.
- AE (Artificial Evolution)
- AL (Artificial Life)
- EC (Evolutionary Computing)
- EC Works when you have a task which has very well defined criteria for a solution
- Simple tasks that are well-defined and have an explicit solution/function
- Artificial Life Forms
- We have a problem finding a definition even
- In the scenario, a high level fitness function develops higher complexity, aka the right-hand mouse seeking pattern, e.t.c. which is never defined
- problem arises when you are trying to synthesize the actual processes which lead to the outcomes
- EC Works when you have a task which has very well defined criteria for a solution
Objective Function-Free Systems
- Systems of self organizing rules or fundamental elements
- Cellular automata
- Taking information from the previous line
- 00000000000
- 00000000000
- 00000010000
- 00000100000
- 00001000000
- Conway's Game of Life
- Problems
- Computation power with computational intractable in the extreme
- We can't define what we want
- Nature's laws are too difficult to model directly in order to "get" life
- It would be extremely difficult to model every particle independently to "simulate" life e.t.c.
- Trying to simulate our universe
- Computational power isn't here and might not ever be here, maybe
- We don't know whether the laws of the universe are changing (annealing)
- 20 years we will have 1,000,000 times as much computing power
- Best possible fixes are to develop a set of rules which are hyper-inclined toward dynamic complexity.
- Require the system to simulate a known system initially and then remove the restriction
- Aggregate interaction could be dependent on low level structure
- e.g. particle interactions results in thermodynamics, but the macroscopic simulation is an oversimplification of the microscopic simuluation
Fastest way to simulate a part of the universe is just to build the part of the universe (can the simulation run faster than itself?)
Disconnect between the intuitions of the people and the system.
They are essentially building a program that does the thing and ensures the success. A true neural network piece of software shouldn't be limited by some fitness function which defines more than it should.
Essentially you get something which looks like innovation, but it is capped to solving the problem one specified. You get the "best" solution to the problem you specified given the constraints.
February 23rd, 2008 at 5:53 pm
[...] http://hplusclub.com/tucson/meeting20080307 [...]