Tuesday, 17 May 2011

Tower of optimisation

In Daniel Dennett's 1995 book "Darwin's Dangerous Idea" he describes a concept he calls the Tower of Generate-and-Test. This is a kind of model of the evolution of intelligent agents. A summary of the tower, starting from the bottom:
  • Darwinian creatures - use natural selection as the generate and test mechanism.
  • Skinnerian creatures - can learn by simple reinforcement learning.
  • Popperian creatures - have a world model, can virtualize sense data and test actions under simulation.
  • Gregorian creatures - tool makers including language and culture.
  • Scientific creatures - Dennett proposes that the scientific method warrants a further floor of the tower.

The model has been extended by Alan Winfield to include a final category:

  • Walterian creatures - artificial, engineered creatures, named after W. Grey Walter, robot inventor.

I think this addition is fairly reasonable. I do think this tower needs to get back to its roots a bit more, though!

If you consider the class of all optimisation processes, you get a rather different picture. Describing each stage in terms of what it adds to the previous level of the tower - and starting from the bottom:

  • Stateless search - optimization with no memory - e.g.: random search.
  • Serial search - uses one agent - e.g.: Newton-Raphson.
  • Parallel search - uses multiple agents - e.g.: simulated-annealing.
  • Splitting - uses agents that can divide and reproduce - e.g.: a simple asexual genetic algorithm.
  • Merging - uses agents that can merge together - e.g. sexual reproduction and parasitism.
  • Learning - uses evolving agents that can additionally learn.
  • Virtualisation - uses agents that can perform most evaluations under simulation.
  • Culture - uses agents that have developed cultural transmission.
  • Artefact symbiosis - the memes start to build tools, minds and bodies for themselves.
  • Genetic engineering - the agents apply their tools to their own germ-line.
  • All-engineered - uses entirely engineered agents - e.g.: machine intelligence and robots.
At the lowest level, the tower is neatly defined - but higher up, it is more like identifying what the major transitions in evolution are. Various steps which seemed fairly important in our world could be interspersed - for example communal living - and writing. For our ancestors, communal living allowed specialization to get going, and writing also provided a major boost. However, these stages do not really seem critical or universal enough to put in this list.

No comments:

Post a Comment