Evolutionary Robots with Fast Adaptive Behaviour in New Environments

This paper is concerned with adaptation capabilities of evolved neural controllers. A method consisting of encoding a set of local adaptation rules that synapses obey while the robot freely moves in the environment [6] is compared to a standard xed-weight network. In the experiments presented here, the performance of the robot is measured in environments that are di erent in signi cant ways from those used during evolution. The results show that evolutionary adaptive controllers can adapt to environmental changes that involve new sensory characteristics (including transfers from simulation to reality) and new spatial relationships. 1 Evolution and Adaptation Evolutionary algorithms are widely used in autonomous robotics in order to solve a large variety of tasks in several kind of environments. However, evolved controllers become well adapted to environmental conditions used during evolution, but often do not perform well when conditions are changed. Under these circumstances, it is necessary to carry on the evolutionary process, but this might take long time. Combination of evolution and learning has been shown to be a viable solution to this problem by providing richer adaptive dynamics [1] than in the case where parameters are entirely genetically-determined. A review of the work combining evolution and learning for sensory-motor controllers can be found in [5, 9]. Instead of simply combining o -the-shelf evolutionary and learning algorithms, in previous work we presented an approach capable of generating adaptive neural controllers by evolving a set of simple adaptation rules [6]. The method consists of encoding on the genotype a set of modi cation rules that perform Hebbian synaptic changes [2{4] through the whole individual's life. The results showed that evolution of adaptive individuals generated viable controllers in much less generations and that these individuals displayed more performant behaviors than genetically-determined individuals. In this paper, we describe two new sets of experiments conceived to measure the adaptation capabilities of this approach in environments that are di erent from those used during evolution. The results are compared to standard evolution of synaptic weights and to evolution of noisy synaptic weights (control condition). Fig. 1. A mobile robot Khepera gains tness by nding as fast as possible the stick under the oor. The walls are covered with white paper and the oor is transparent. The robot has a sensor pointing downwards that can detect the stick when it passes over it. The stick can be positioned at any location under the oor. The sources of change address two major aspects of behavioral robustness: sensory appearance and spatial relationships of key-features of the environment. 2 Experiment I: Changing Sensory Appearances A mobile robot Khepera is positioned in the rectangular environment shown in gure 1. The walls are covered with paper and the oor, which is transparent, is placed on four supports. A stick is positioned at a random location under the oor. Each individual of the population is tested on the same robot, one at a time, for a maximum of 500 sensory motor cycles, each cycle lasting 100 ms. At the beginning of an individual's life, the robot and the stick are positioned at random positions. The tness function selects individuals capable of nding the stick in the shortest time, = 1 t