If left long enough, an evolutionary algorithm will usually converge to a single solution. This is not always desirable; if the EA is optimising a multi-modal problem, the user may want to know the location of more than one optimum. Speciation1) encourages the EA to do this by dividing the population into species, usually by clustering individuals that are close to each other in the decision space. Speciation also reduces the risk of premature convergence - even if one species has converged on a local optimum, there is still a chance another species will locate a global optimum.
By forbidding or limiting interaction between individuals in different species, each species will generally converge on its own optimum2). At the end of the run, the best individual of each species is considered to represent the location of the species’ optimum.