It is difficult to define a set of rules for a cellular automaton (CA) such that creatures with life-like properties (stability and dynamic behaviour, reproducton and self-repair) can be grown from a large number of initial configurations. This work describes an evolutionary framework for the search of a CA with these properties. Instead of encoding them directly into the fitness function, we propose one, which maximises the variance of entropy across the CA grid. This fitness function promotes the existence of areas on the verge of chaos, where life is expected to thrive. The results are reported for the case of CA in which cells are in one of four possible states. We also describe a mechanism for fitness sharing that successfully speeds up the genetic search, both in terms of number of generations and CPU time.