Waterstriders orientate toward the source of surface vibrations by discrete rotational movements. We describe here an artificial neural network that simulates this behaviour and compare the results to published data (Murphey 1971). All simulations were performed using the Macintosh version of 4.05 Neural Works Professional II Plus. A backpropagation network was used with a learning coefficient of 0.10, momentum of 0.40, sigmoidal transfer function, six input units, one corresponding to each leg (vibration receptor) of the waterstrider, and an output corresponding to the elicited angle of rotation. The waterstrider was situated at the centre of a circle upon whose perimeter vibrations were taken to originate. Distances from the origin of a vibration to each leg comprised the input data. The network was trained with a full compliment of legs to rotate towards the source of vibration. It was subsequently tested with all receptors present and a variety of legs amputated (combinations of 1, 2, or 3 removed from the right side). To simulate an amputation the appropriate input unit's connections with the output were severed. Training and test sets consisted of, respectively, 400 and 200 vibrations randomly spaced on the perimeter of the circle.

The network reached convergence during training in approximately 1200 iterations. When tested with all receptors present a linear relationship (r² = 0.9999) was found to exist between the desired and obtained rotational angles. Amputation of one or two receptors resulted in marked deviation from linearity within the angular range of detection corresponding to that of the amputated receptor(s), while amputation of three receptors resulted in the network rotating contralaterally to all vibrations originating ipsolaterally to the amputated side. All trials produced results that corresponded qualitatively to Murphey's (1971) data indicating rotation is largely dependent on the first remaining receptor to be stimulated by a vibration. We conclude that an extremely simple artificial neural network is capable of qualitatively simulating the orientation of behaviour of waterstriders to surface vibrations.

R.K. Murphey (1971) Sensory aspects of the control of orientation to prey by the waterstrider, Gerris remigis. Z. vergl. Physiol. 72: 168-185.