Seminar Lecture of Micky Frankel (025528837).
 

• Abstract
• Introduction
• AI versus AL approach to Cognition
• The AL modeling approach
• Exemplifying Animal Intelligence
• Models for Synthesizing Animal Intelligence via Evolution and Learning
• Research Issues
• Methodological Principles

Introduction

A major goal of Artificial Life research is to gain insight into both life as it is and life as it might have been. Here we focus in understanding the nature of intelligence from an AL perspective, that is, the evolution and development of complex nervous systems supporting cooperative behaviors. We interest in the way artificial neural networks support cognitive processes and in the way intelligence is distributed within groups or populations of individuals, with a special focus on the role of communication in survival strategies requiring cooperation.
 

1. Environments -

simulates worlds whose conditions match, at some level of abstraction, those selection pressures in which a variety of animal behaviors may evolve or develop.
 
2. Processes of genetic expression -

mapping from artificial genomes to phenotypes that control behavior. A genotype is commonly a bit string; phenotypes is often some type of artificial neural network or connectionist architecture that controls the artificial organism’s behavior, through the simulation of sensory/motor neurons and interneurons.
 
3. Learning and development -

methods under genetic control for modifying or growing the nervous systems of artificial animals during their lifetimes.
 
4. Evolution -

recombination and mutation of parental genomes during mating to produce variation in their offspring.

Models for Synthesizing Animal Intelligence via Evolution and Learning
 

Evolution and Learning of Food Discrimination

Todd and Miller  set up an abstract, simulated "aquatic" environment containing two distinct patches of "plant material" - a red patch and a green patch. Within one patch the red plant served as "food" for the evolving creatures, while the green plants act as "poison". In the other patch the color roles are reversed. In both patches "food" has a "sweet" smell while "poison" has a "sour" one. The "turbulence" in the water determines how accurate the smell is. ("Food" may mistakenly smell "sour" and vise versa). Each creature remains during its lifetime, in a given patch. However its offspring may be placed at birth in the other patch. If a creature eats "food" its metabolism increased thus improving its reproductive success, "poison" consuming, however, reduce metabolism.

A neural network with learning capabilities controls the creatures. Over several hundreds of generation a creature was evolved with a hardwire connection between smell and the eating motor neurons, and a learnable connection between color and motor neurons. This connection was modified over the lifetime of a given creature.

Evolution of Foraging and Trail Laying

Collins performed a series of experiments in which he attempted to evolve colonies of artificial neuron network ants that both forage for food and lay down pheromone trail to guide other to food sites. In earlier experiments food foraging behaviors evolved, but they were non-ants like, for examples they walked in circles or semi-circles. Only in later experiments Collins succeeded in evolving ant like behavior, these ants walked mainly forward with random turns until food was found, then they transport food back to the nest while laying a trail of pheromones. Collins forced generations 1000 to 2000 to involuntarily lay trails and then returned the pheromone release control back to each ant. At generation 2001 there was a large decrease of the amount of pheromones released, but the ants evolved to both lay and follow these trails by generation 2100. Collins theorized that before ants were forced to release pheromones, trail following could not evolve thus trail laying could not evolve either. However, once ants evolve to follow trail, when trail laying was mandatory, trail following could evolve.
 

Evolution of Communication

MacLennan made an experiment in which simorgs share an environment where they can match and post signals. Whenever a simorg's action matches that of the most recent symbol posted, both the sender and the receiver receive a credit. Simorgs with higher credit have better chance of reproduction. MacLennan's experiment showed that enabling communication and learning result higher average fitness.

In Werner and Dyer simple communication protocols for mating were evolved. In this model male and female were ordered on a two dimensional grid where male are blind but mobile, while female can see (that is sense the existence of a male in its surroundings), yet immobile. Male and female can both send signals and receive them. A couple of male and female can mate only when the male lands on a cell with a female. A communication was evolved where when a male got near a female; the female directed him towards her, by signaling the direction for movement, while the male was evolved to follow these directions. In addition sub-species with different signaling protocols ("dialects") evolved and competed in the environments. When partially permeable barriers were set on the grid, separate sub-species were evolved and survived, in spite of occasional migration and contact from member of other sub-species.

Evolution of Predation and Predator Avoidance

Werner and Dyer extended their two-dimensional model with simulation of smell and sound, and with multiple of species interact on the grid. The new environment contained objects as trees, plats and holes. Each organism involuntarily signals his "smell", an information about his species, gender etc. As a creature moved faster, the louder sound it makes thus can be heard from mote distant place. A sound in a variety of "frequencies" can also be produced voluntarily, that is under the neuron control. In one experiment herbivore plants eating dog were created, using as prairie to snakes (than cannot climb trees) and hawks (which cannot get into holes). The dogs were evolved to run away from snakes and hawks, while the latter evolved to chase the dogs. The dogs were also evolved to form hers to protection from predators, they also evolved different predator warning signals, enabling the receiver dog to seek appropriate shelter based on the nature of the warning signal.

I feel that one cannot gain much knowledge out of these experiments, for most of the results were dictated by the conditions of the experiments. The experiment Todd and Miller  was designed in such a way were the only stable solution (except extinction) and the best one is evolution of smell distinction and learning of color distinction. The same goes with MacLennan's experiment were communication evolving was the solution the experiment was directing into and with the first experiment of Werner and Dyer where guidance from the female to the male is the only solution (except random behavior). In their second experiments enabling the dogs to communicate and designing different shelter for different predator makes the result of this experiment quite expected. In Collins experiments the conditions were compelled (brutally in my opinion) to force de desired result.

I think that these experiments exhibit the power of biological computation in solving problems of finding the solution with the highest grading, but the nature of this solution does not have biological implication, since it was design along with the experiments.

There are two exceptions; the facts that the dogs in Werner and Dyer environments tend to group in herds and the different dialects which were evolved in their other experiment. Nevertheless one should be extremely cautious with zoological conclusion from these facts, since the simulation of the real world is quite degenerate. And there is no evidence that the simulating of the brain neuron network is similar to the artificial one in a level higher that the physical one, that is coding information in the model is not necessarily resembling to the one in nature.

• The relationship between learning and evolution
  
  
• The ways of coding knowledge into artificial neuron network and of altering the network (learning) as respond to the environment.
  
  
• Methods for phenotypes encoding into genomes
  
  
• Addressing outstanding problems in population, as evolution and altruism
  
  
• Encoding of environments that are computationally tractable but rich enough to supply the necessary selectional pressure.

• Principle of Abstraction Hierarchies -

The level of abstraction should adapt the objects of the experiments. Simulating a nest building by firing a nest-neuron can suffice in a model that experiment the placement of the nest but bot one that experiments evolution of nest building behavior.
 
• Principle of Minimal Effective Embodiment

It is important to focus on the evolution of the structure of the nervous system while avoiding, where ever possible, the inclusion of physical properties. In a prairie and predator experiments allowing alternation of the organism size may lead to a race where each species is growing in each generation. Whenever experimenting for hypothesis of physical attributes, it is suggested to slow the physical change, to allow neural adaptation.
 
• Principle of Midpoint Entry -

Because the search space is so vast and broad range, it is acceptable to start one's models somewhere in the middle. This principle supports a mix of top-down and bottom-up strategies. This mixed approach allows one to start with a population with pre-engineered capabilities, and then see if these capabilities can maintain or even improve.
 
• Principle of Indirectness -

Whenever possible, it is important to avoid setting up an objective function that select directly the desired behavior. Instead the desired behavior should arise indirectly as a side affect of the increased survivability.
 
• Principle of Naturalness -

Because there are potentially so many parameters to adjust in AL systems, it is important to impose constraints that are intuitive. A major source of intuitions comes from constraints that already exist for real animals in nature. Whenever these intuitions are violated, they should carefully noted.