Evaluation safety for real functions

[morinim]

Evaluation safety is a component of the closure property.

It's necessitated by the fact that many commonly used functions can fail in various ways (divide by 0, square root of a negative number...).

This is typically dealt with by appropriately modifying the standard behavior of primitives. It's common, for example, to use protected versions of numeric functions that can throw exceptions, such as division, logarithm, and square root. The protected version of such a function first tests for potential problems with its input(s) before executing the corresponding instruction, and if a problem is spotted some pre-fixed value is returned.

At the moment we're using an alternative to protected real functions: we trap NaN results and strongly reduce the fitness of programs that generate such values. This method can lead to all the individuals in the population having nearly the same (very poor) fitness, leaving selection with very little discriminatory power.
On the other hand it can cut the evaluation time of poor individuals.

Further experiments are required.

[morinim]

Also the current approach is slightly different compared to Vita.

+inf / - inf values are considered valid and used as intermediate results and final output.

Previously (Vita) they were treated the same as NaN.

[morinim]

At the moment we trap both NaN and infinity values.

Measurements don't show performance variations and the source code is simpler.