Evolutionary programming could be a path to artificial general intelligence if one can find a way to increase it's complexity tolerance.

In evolutionary programming much has been done to improve the evolutionary algorithms but not so much to make the substrate language less brittle. I am not aware of theoretical grounds that would indicate the impossibility of developing a low brittleness programming language (LBPL) that would enable evolution of complex code. Humans are proof that there exists some mechanism for evolving complex code considering, for example, that a carpenter can build a house and a programmer can write a large program - provided one accepts that brains are Turing-Adequate, by which I mean not more than Turning-Complete. 

To illustrate approaching a LBPL consider a conventional programming language construct:

if (A and B) or (C and not B): then do something

The chance of a mutation to this line of code being successful is low compared to the following approach. Suppose we revise the programming language so that all conditionals are done via truth tables. In the above example we have:

A	B	C
0	0	0	0
0	0	1	1
0	1	0	0
0	1	1	0
1	0	0	0
1	0	1	1
1	1	0	1
1	1	1	1

Now we can mutate the single bits of the rightmost column with a 50% chance of a successful mutation and evolve an optimal table in 8 mutations. The granularity of mutation has been lowered to one-bit, thereby increasing the probability of successful mutation exponentially. In contrast, trying random formulas would require a far larger number of permutations.

As a second example, consider evolution of a formula such as Y = X ^ 2.

One can divide the plane into a coarse grid and then do bit mutations on each cell, testing the performance of the function iteratively. After reaching a level of stabilization the grid resolution could be increased.

Using language structure re-engineering it may be possible to devise an effective LBPL.

Posted 8/5/07, Revised 8/14/07, 0/22/07