Hi, many thanks for the replies. I also am inclined to agree that this is
not a bug with ecj. I too get much worse results for symbolic regression
especially, and also when using the jGE software package, so this is not
just ecj. I have looked quite deeply into the ecj code, and the only
difference I can see is that there is no GE 'wrapping' operator, where the
mapping process goes back to the first codon if there are not enough in the
genome. I have written a hack for GPSpecies which includes this, and yet the
results are only marginally better. Was the wrapping operator left out
intentionally? The countNumberOfChromosomesUsed array seems ideal to count
the wraps, and yet it is not fully used, and currently only has a size of 1. 
Just to clarify, in case anyone reads this in future, I am indeed using the
Santa Fe example code, provided with ecj. I have set the number of steps to
be 615 rather than 400, as I think that number appears in a future paper.
The example needed some modification as the grammar and parameters were not
the same as in the paper.