Saturday, May 18, 2013

Evolutionary criteria and goal for intelligence

2013
Capability to think at higher level of abstraction and build longer inference chains as matter of physical dominance on the evolutionary path due to lower consumption of energy and lower reaction time

Following the previous analysis, spawned by Ben Goertzel's definition of intelligence as "the ability to achieve complex goals in complex environments, using limited resources" I was thinking about the goals for the propsed agents of distributed intelligence, given the resourece and environments are well defined.
And it came down to the goals of intelligence properities of live being required for survival on the evolutionary path.
Frst, let's say we have two intelligent animals exposed simple patterns food patters such as XYA, XYB, XYC, and XYD. Assume animal ONE is dumb enough to allocate neural cells required to keep all four patterns so once any of them is exposed, the food is recongnized. Assume animal TWO can perfrom high-level abstraction thinking so it can reason out that there is only one food-relevant pattern XY so only the one worth allocating neurons for it. Thus, animal TWO needs at least 4 times less neurons fed to provide the same amount of food. Respectively, the animal TWO has at least 4 times less probability to die from hunger (speaking speculatively) or it can allocate remaining neurons to say recognize dangers, increacing its chances to survive the other way around.
Which suggests that capability to think at higher level of abstraction assures competitive advantage as matter of physical dominance on the evolutionary path due to lower consumption of energy.
Next, say these two animals exposed the danger condition which can be recognized as sertain sequence of events such as P Q R S. Assume animal ONE always checks the whole sequence to to be experienced before activating escape strategy. Assume animal TWO is capable to infer appearance of P implies appearance of  S eventually, so it activates the escape strategy at the point of Q. Obviously, due to the reaction time 4 times shorter for the latter agent, chances for survival are forth larger than in case of the former agent. Which suggests that capability to build longer inference chains assures competitive advantage as matter of physical dominance on the evolutionary path due to lower reaction time.

This can be transposed to our coming universe of software agents.
Say we have two agents given the same amount of memory both need to recognize some event on the stock market in order to launch the trade. In case of agets operating with paged memory or cached graph database, agent which need 4 times less memory to recognize the pattern, has 4 times less probability the required memory is not out of the cache or memory page, and so it has more chances to launch the trade. At the same time, the owner of the agent has to spend 4 times less money to buy memory in order to operate the business. Same example applies to reaction time where agent capable to infer consequences from early signals gives more opportunities to act first rather than agent waiting for the exposure of event chain in full.

Which suggests that one formal criteria for comparative intelligence of a black-box software agents can be stated as follows. Given same set of inputs associated with respective set of outputs, after the course of training be able to perform the task consuming less memory and in lesser time.

Further, it is expected (to be proved further), that extending the task set to the extent some tasks would be failed due to complexity, the more intelligent agent (accordingly to the former test) would fail less tasks in general, at least due to extra free memory available and better general intelligence.

Back to the biological evolution, this may turn into answer for why dolhins and elephants are not smarter than humans even having more neurons - because they don't reason at such high levels of abstraction and don't take time to shorten reasoning chains, wasting much of these extra neurons unefficiently.

How this all refers to connectivity and topology is another story...