To develop robust agents capable of adapting to changes in environment conditions we analyze the effects of introducing noise into multi-agent communication and varying prey movement patterns in a version of the predator-prey pursuit problem. In this simulation, predators communicate with languages evolved using a genetic algorithm. We show the time it takes to capture prey increases as more noise is introduced into the predators' communication string. We show that predator performance under noisy conditions can be improved by training the predators with noise. Increasing the amount of noise during predator training increases the performance capability of the predator, thus reducing the amount of time it takes to capture a prey in noisy conditions. Training against different prey movement patterns can also enhance predator performance. We show that predators trained with prey that stop and move randomly perform best when tested against various prey movements. Alternatively, predators trained against constantly idle, or frozen, prey showed the worst performance. When any type of predator was tested against the frozen prey, the predators had some problems completing the capture. Drawing inspiration from the popular film Jurassic Park, we refer to this phenomenon as the T-Rex effect. Prey that hide or do not move have the greatest probability of surviving. Training predators with factors that normally decrease testing performance helps develop robust predators capable of increased performance under both adverse testing conditions.
展开▼
