An evolutionary approach to autonomous agent design: investigating the interdependence between morphology and control
We employ methods inspired by biological evolution and development to investigate the importance of morphology for intelligence. We apply artificial evolution to the design of complete agents, including both control architectures (artificial neural networks) and morphology (sensors, actuators, and body shape).We use an artificial evolutionary system (AES) to evolve simulated agents that can complete some specific task. Particular attention will be devoted to the role of the morphology of these robots with regard to their fitness in a specific environment. These simulated agents are then used as blueprints to build real world robots with our flexible robot building kit.
The evolutionary cycle:
(click on the thumbnails to get animated gifs)
1st
step: We evolve artificial creatures with different morphologies
using Peter's
AES.
The
interconnections of neurons are also determined by the morphology.
2nd step: The artificial organism is tested in a simulator. Creatures with very low fitness are removed.
3rd
step: Real robots are built according to the blueprints of the
artificial organisms that survived step 2. These robots are then tested
in a real world environment.
The fitness values of the real robots are used to select mating partners and create offspring using some evolutionary strategy. Then the evolutionary cycle recommences at step 1.
An example: Whiskers vs. Light Proximity Sensors
(click on the thumbnails to get animated gifs)This robot evolved two whiskers on its front. It seems from first experiments that it gets stuck less frequently than an earlier organism using only proximity sensors based on light reflection.
