Evolutionary Robotics
While most work of ER focuses on technical issues concerning the evolution of neural controllers for wheeled robots (in the line of Floreano and Nolfi), there are also efforts in the evolution of morphology and behavior of physically simulated walking/crawling creatures (following Sims). Both branches of ER can be modelled in Framsticks.
Virtual life projects in this wide field can be grouped in three themes:
- Evolution of Morphology and Behavior
- Robot and Animal Navigation
- Interaction between Learning and Evolution
1. Evolution of Morphology and Behavior
Sims' (1994) evolved virtual creatures offers an inspiring visual example of 'life-as-it-could-be'. As one of the first simulations of evolution of bodies and controllers of creatures in 3D physical simulation, it can be regarded one of the ancestors of Framsticks.
Although evolving various morphologies and behaviors is easily done in Framsticks, this is not the aim in Virtual life projects. There are many interesting possible research projects that go beyond standard optimisation.
Some possible projects include:
- Evolution of morphology by natural selection
Most experiments in evolving morphology use explicit fitness functions, by assigning fitness based on displayed velocity and selecting the speediest for reproduction. Even the competitive co-evolutionary experiment conducted by Sims used explicit fitness function, albeit one in which the fitness of one creature was co-determined by the performance of the other.
Evolving creature morphologies in artificial ecosystems with endogenous fitness (i.e. without explicit fitness) remains, however, largely unexplored territory. Such simulations include populations of co-evolving creatures that compete for energy sources and reproductive resources. Since many morphological traits (though certainly not the peacock's tail) are the products of the struggle for survival, the evolution of morphology seems best modelled by endogenous evolution.
See the description above. In this project, however, not only the struggle for survival is modelled, by also the struggle for reproduction. Some morphological traits, such as the peacock's tail, are not the product of natural selection, and are even a disadvantage for survival, but are products of sexual selection.
Suggested reading:
2. Robot and Animal Navigation
The most basic behavior an organism can engage in is moving through its environment. Even with the simplest of organisms, this is not only a passive process (through current, wind, gravity), but also involves active goal-directed movement using some sort of navigation mechanisms. Braitenberg (1986), and John Hammond's Seleno (1915) long before him, has illustrated how this can be easily done mechanically and how easily we, human being, are (mis)interpreting the simple taxic behaviors as rational goal-directed behavior.
Still, it is hard to see how more complex behavior -e.g. anticipatory, decisive, sequential- can emerge from Braitenberg-style creatures without use of symbolic representations. For example, how can creatures learn to find food in a complex maze without using some cognitive map of your surroundings? Or how to find your way back home, without a good clear map?
- Comparison between evolved neural controllers behaviour and Braitenberg architecture
- Spiking Neural networks: biologically plausible time-dependent neural networks for robots (Floreano, van Leeuwen)
Khepera-like creature in Framsticks 
Because many experiments in evolutionary robotics (ER) use Khepera robots, a Khepera-inspired creature is developed in Framsticks to allow replicating and extending projects reported in existing literature. This (simulated) robot forms a platform that offers a wide variety of possibilities for experiments in evolution of behavioral tasks, neural networks and evolutionary learning mechanisms.
The creature is modelled in Framsticks using the f1 genotype encoding like X(X,X,X,X,X,X,X).
On two opposing
sides, two 'wheel neurons' are attached. The activation of the left wheel causes the creature to rotate around the right wheel (and visa versa). Sensors can be arranged on the endings of the sticks.
Download this creature
3. Baldwin effect: Interaction between Learning and Evolution
ALife modeling throws a new light upon evolutionary phenomena. An excellent example is investigations of the Baldwin effect. According to the Baldwin effect, the learned features of organisms could be indirectly inhered in subsequent generations. The Baldwin effect works in two steps. At the first step evolving organisms obtain (through appropriate mutations) an ability to learn a certain advantageous trait. The fitness of such organisms is increased; hence they are spread throughout the population. But learning is typically costly for an individual, because it requires energy and time. Therefore the second step (which is called the genetic assimilation) is possible: the advantageous trait can be “reinvented” by the genetic evolution and become directly genetically encoded. The second step takes a number of generations; a stable environment and a high correlation between genotype and phenotype facilitate this step. Thus, the advantageous trait that has been originally acquired can become inherited, though the evolution is of Darwinian type. (from Principia Cybernetica Web)
There are many ER experiments in this line of research. Only a brief summary of projects is given with links to websites and articles to allow further reading.
- Evolutionary Reinforcement Learning
- Evolving neural network learning rules for robots
- Floreano, Urzelai (2001) Evolutionary Robotics: the Next Generation, proceedings of Robo Sapiens, Utrecht, April 2001.
|
