GE, explosive grammars and the lasting legacy of bad initialisation

This paper explores some of the initialisation schemes that can be used to create the starting population of a Grammatical Evolution (GE) run. It investigates why two typical initialisation schemes (random bit and ramped half and half) produce very different, but in each case skewed, tree types. A third methodology, Sean Luke's Probabilistic Tree-Creation version 2 (PTC2), is also examined and is shown to produce a wider variety of trees. Two experiments on different problem sets are carried out and it is shown that for each of these test cases, where the “wrong” initialisation method is utilised, the chance of achieving a successful run is decreased even if the runs are continued long enough for the populations to stagnate. This would seem to suggest that the system does not typically recover from a “bad” start.