Both sets of experiments (i.e. on the Santa Fe and on the random trails) runs were conducted with a range of fitness plagiarism penalties. The plagiarism penalty is applied to programs which, when run on the same test as their first parent, have the same score as that parent. (The first parent is defined to be the one from which they inherit their root node). E.g. when using the Santa Fe trail, fitness is reduced if a program causes the ant to eat the same number of food pellets as the program's first parent did. In the case of the dynamic test case, the program must be run both on its own test and the test for the previous generation, i.e. the test its parents where run on.
The smallest penalty (referred to as -0.5) causes the fitness to be reduced by half a food pellet. The other penalties reduce a program's fitness by a fixed fraction. The highest penalty (100%) sets the fitness to zero. As tournament selection is used, the effect the penalty has on the number of children a program has is complicated, however when the penalty is large compared to the spread of scores within the population, even the best program has little chance of producing children for the next generation.
In many of the results that follow the data falls into a low penalty group (20% or less) and a high penalty group (50% or more). We can estimate the mean plagiarism penalty (when applied) as its size times the mean score in the population. The separation between runs where the penalty is effective and the others corresponds to when this estimate is bigger than the variation in program score across the population (as measured by its standard deviation). That is the plagiarism penalty has little effect if it smaller than the existing variation in programs' scores before it is applied.