We developed a virtual world to study the effect of predators on predator recognition. We trained a neural network to discriminate between the shapes of simulated aerial and terrestrial predators and non-predators. Then, the network's weighting values were fixed into the genomes of a set of autonomous agents. These animats were required to eat, avoid death due to starvation, and avoid predation, by fleeing from approaching predators. We systematically varied the predator's lethality, the mutation rate, the cost of fleeing a predator, and the presence or absence of aerial and terrestrial predators. We used ANOVA to analyse the average recognition ability (a measure of directional selection) and the standard deviation of recognition ability (a measure of relaxed selection) after 500 generations of selection. Mutation rate and the cost of flight had the greatest effect on both the average and standard deviation of recognition abilities. The loss of all predators relaxed selection on predator recognition abilities. The loss of specific predators had complex effects on recognition abilities. Persistence is largely influenced by escape costs.