Two years ago, a UK-based company, Cyberlife, unleashed a game called Creatures onto the PC platform. When both Oxford geneticist Richard Dawkins and Hitchhiker’s Guide to the Galaxy author Douglas Adams praised it as a scientific and artistic masterpiece, it helped me justify the amount of my time it was consuming. This was "real" artificial life: characters called Norns that had neural networks for brains, were capable of interacting with their peers and the user, and used a genetic algorithm and "digital DNA" to allow for mating and reproduction. They looked like a cross between "a deer, a cat and a hamster" and were, well, pretty cute. You taught them to eat when they were hungry, to sleep when they were tired, and to avoid all sorts of peril in their two-dimensionally rendered little world. As crazy as it may sound, if I had to pick the single most significant influence on my decision to study Synthetic Characters at MIT’s Media Lab next year, it would have to be Creatures.

Creatures raised more questions than it answered. Because it is a commercial product, its creators are deliberately sketchy about what makes their creations tick. Dr Bruce Blumberg, head of the Synthetic Characters group at the Media Lab, suggested to me that there’s probably "less to Creatures than meets the eye." That’s probably true, but it doesn’t take away from what might evolve from products like Creatures.

I was interested to find Stefano and Parisi’s article in the Handbook of Brain Theory and Neural Networks that addressed the issues of how to encode the "genome" of an artificial creature. They introduce their article by discussing the elements we might find in the genotype of a neural network. These elements might include any or all of the following:

• The network architecture
• The initial weights
• The learning rates and momentums

To create a "genotype," we need to somehow encode this information into a string that can be passed between generations. In so doing, we tend to depart from the biological underpinnings. The architecture of a neural network is typically encoded by specifying which neurons connect to which other neurons. This is a far cry from a three-dimensional physical reality, where neurons close to each other in 3D space are more likely (but not guaranteed) to form connections.

Nolfi and Parisi propose a new technique that more closely models the brain’s development. They create organisms they call "Os" that search for food in an artificial environment. Their life-span is finite and pre-ordained (which isn’t very lifelike, really), and the one-fifth of the organisms that eat the most food over their life-span are allowed to reproduce agametically, which means they simply generate five copies of their genotype. Random mutations are introduced into the copying process by replacing 20 of their 40 genotype bits with randomly selected values.

Each of the 40-genotype blocks has five components that encode the following information:

• Whether or not the neuron exists at all
• The (x, y) co-ordinates of the neuron in 2-D space
• Branching angle of the axon
• Segment length of the axon
• Activation bias or threshold of the neuron

Neurons were grouped into three groups: sensory (5), motor (5) and internal (30).

The first incarnation of the Os didn’t learn at all during their lifetimes. Their behaviour was entirely genetically transmitted. This differs from the Norns of Creatures, who learn from the user and their peers throughout their lives. Nolfi and Parisi sought to discover how learning during a creature’s lifetime would affect the evolutionary process.

To do so, they split the network into two parts: the "standard" network, and a "teaching" network. Both networks shared the same input units, but had separate sets of internal and output units. The standard output still generated motor actions, and the teaching units were equipped with pre-ordained network parameters that were considered to be "good." Using a back-propagation algorithm with a learning rate of 0.15, the "teaching" network schooled the standard one. With this modification, evolution tended to favour Os that had a predisposition to learn, and not the ones that started their lives in the best shape for acquiring food. Interestingly, their performance at birth did not increase across generations; instead, what increased was the ability to learn the desired performance.

Nolfi and Parisi may have come closer to a genotypical model that takes into account the location of neurons in a three-dimensional space, but they’ve also strayed a few steps further away from a "real life" model:

• Forcing creatures to each live the same lifespan is unreasonable, particularly when their performance is based on their ability to acquire presumably life-sustaining "food."
• The tremendous amount of mutation that happens between each generation is also unrealistic. Whereas far less than 1% of the genes in a mammal undergo a mutation between generations, here a full 50% are mutated.
• We are given no indication of how they arrived at the pre-ordained "teaching network." Living organisms don’t have the equivalent of a "Schaum’s Outline" guiding their evolution! Within their genetics, the Os already have bestowed upon them a genotype they are to consider "ideal." Not only does this discourage creativity and push Os towards a single pattern, but a parallel doesn’t exist in the real world.
• The authors concede that they are also missing a model for the mapping between genotype and phenotype (ontogeny), which Cyberlife claims it has already deployed within its Norns.

The results Nolfi and Parisi obtained are interesting, but to generalize their results beyond the realm of their "Os" would be unrealistic. The notion that "ability to learn" is more important than "what you are already encoded with" is an appealing one, but I certainly don’t think any correlations could be drawn between these artificial creatures and traditional living beings, or even the Cyberlife Creatures.

Cyberlife’s creatures suffer from similar limitations:

• The Norn brains intrinsically group objects into categories like "food," "toys" and "lifting devices." Regardless of how the user tries to train them, they are forced to associate objects with one (and only one) of these groups. So, you can teach a Norn to eat "food" when it is hungry, but you can’t teach it to eat carrots and not cake.
• The Creatures don’t have an "ideal" genotype like the Os, but they are provided with a "computer console" within the game is can be used to teach the Creatures a set of about a dozen verbs. These verbs are also pre-programmed into the system. Norns have no way except by using the computer to learn the meaning of a verb, which is perhaps as much a fault of the interface as anything else. It would certainly be difficult in a point-and-click environment to teach the meaning of the word "eat," for example.

Both Norns and Os provide exciting starts towards artificial life. It is clear, however, that there is a long way to go. As for the Norns, it would appear that the rigidity of the input-output relationship in their neural networks is a stumbling block. In Nolfi and Parisi’s case, they take one step towards more authentic artificial life, and three steps away from it. Perhaps genetic algorithms and neural networks are not the only tools we will need to use if we wish to create authentic artificial life.

Cyberlife Web site, http://www.cyberlife.co.uk.

Nolfi, Stefano and Domenico Parisi, "‘Genotypes’ for Neural Networks" in The Handbook of Brain Theory and Neural Networks (Michael A. Arbib, Ed.) pp 431-434.

Haykin, Simon, Neural Networks: A Comprehensive Foundation (1999) 2^nd Ed.