One of the roads currently being explored to create human-equivalent AI is to mimic biology and evolve it. Artificial evolution is a source of optimism for AI researchers who hope to one day be witness to technological singularity. But if human evolution which has required a colossal amount of energy is any indication, argues electronic engineering professor Alan F.T. Winfield, than energy cost may thwart those aspirations.

In his paper Estimating the Energy Cost of (Artificial) Evolution, professor Winfield does just that: determining how much energy has gone into getting us from single celled organisms to our present form. With this exercise he aims to counter the optimism that artificial evolution is a viable road toward singularity: a radical break in human history as a result of the emergence of superintelligence causing an unprecedented acceleration of science and technology.

Energy capture
Winfield sets out to estimate the minimum and maximum energy cost of human evolution. To establish the maximum value he calculates how much solar energy has been captured on earth through photosynthesis. Using estimates of how much energy is stored in biomass annually and going back some three billion years, Winfield concludes that earth's biomass has approximately captured ∼ 5.7 × 1012  EJ of the sun's energy.

This number is the upper bound of energy cost. Winfield: “This represents an estimate of the total amount of energy available to natural evolution, to date. Furthermore, this was the energy available to evolve all living things that have ever existed, including humans."

Working backwards
Winfield calculates the lower bound by tracing back human ancestry to single celled organisms via hominids, primates, mammals etc. and calculating for each of these evolutionary stages the average daily energy consumption of each individual, the size of the population and the average age of reproduction. He arrives at a minimum energy cost for human evolution of ∼ 8000 EJ. For comparison, the total human energy use in 2010 was about 539 EJ.

Artificial evolution
Winfield then proceeds to calculate the energy cost of artificial evolution. The most complex artificially evolved robot controllers currently available are more or less equivalent to a nematode roundworm with 302 neurons and ~ 5000 synapses. The evolutionary process requires a population of a 100 creatures evolving in an artificial environment for about a 1000 generations. An average computer, writes Winfield, “may complete the task in 10 hours, at a total energy cost of ~ 9000 KJ."
 
Kleibers_Law
Kleiber's Law. Source Wikipedia
 

Complexity
The energy cost for evolving a single virtual creature in this scenario is 9J / hr. However, this is a fairly simple creature. With increasing complexity the energy cost rises. To get an idea of how much it increases Winfield proposes to project Kleiber's Law on the virtual world. Kleiber's Law states that the energy needed by an animal to sustain itself increases linearly with increasing body weight. “Perhaps a similar relationship might exist between, say, neural complexity and energy cost e for virtual creatures”, writes Winfield. If that is the case evolving a human-equivalent AI would be about 1010  - 1012  more energy-intensive than the nematode roundworm's virtual cousin.

Challenging environment
But 'simply' providing the energy for such a process isn't even enough, says Winfield. Complex biological organisms like humans did not evolve in a vacuum. A big part of the evolutionary process is facilitated by having to adapt to complex environments and learning to exploit them. Likewise, complex virtual environments will need to be provided to serve as sparring partners for artificial creatures to boost their evolution. Even when setting aside the difficulty of engineering such environments, the added energy cost would be huge.

Winfield concludes that we might need to reconsider our hopes of creating superintelligence through artificial evolution.
 
Via: robohub.org

Image: futurebuff.com