Evolutionary autonomous agents and the nature of apraxia

Abstract

Background: Evolutionary autonomous agents are robots or robot simulations whose controller is a dynamical neural network and whose evolution occurs autonomously under the guidance of a fitness function without the detailed or explicit direction of an external programmer. They are embodied agents with a simple neural network controller and as such they provide the optimal forum by which sensorimotor interactions in a specified environment can be studied without the computational assumptions inherent in standard neuroscience.

Methods: Evolutionary autonomous agents were evolved that were able to perform identical movements under two different contexts, one which represented an automatic movement and one which had a symbolic context. In an attempt to model the automatic-voluntary dissociation frequently seen in ideomotor apraxia, lesions were introduced into the neural network controllers resulting in a behavioral dissociation with loss of the ability to perform the movement which had a symbolic context and preservation of the simpler, automatic movement.

Results: Analysis of the changes in the hierarchical organization of the networks in the apractic EAAs demonstrated consistent changes in the network dynamics across all agents with loss of longer duration time scales in the network dynamics.

Conclusion: The concepts of determinate motor programs and perceptual representations that are implicit in the present day understanding of ideomotor apraxia are assumptions inherent in the computational understanding of brain function. The strength of the present study using EAAs to model one aspect of ideomotor apraxia is the absence of these assumptions and a grounding of all sensorimotor interactions in an embodied, autonomous agent. The consistency of the hierarchical changes in the network dynamics across all apractic agents demonstrates that this technique is tenable and will be a valuable adjunct to a computational formalism in the understanding of the physical basis of neurological disorders.

Figure 1

(a) Trajectory of a robot which was able to accomplish both tasks; only the last 3 movements are shown. The robot begins at the open triangle. It then moves to position 1 (sustained light) and then to position 2 (brief light flash). (b) Activity from each neuron of the network and from the light sensor; data for the entire epoch is shown which constituted movements to 5 different sustained lights followed by 1 brief light flash. (c) FFT data for all five neurons calculated over the entire epoch; the value of the corresponding Hurst exponent for each neuron is on the top right.

Figure 2

(a) Trajectory of a lesioned robot which lost the ability to move to the brief light flash; only the last 3 successful movements are shown. The robot begins at the open triangle. It then moves to position 1 (sustained light) and then to position 2 (sustained light). The robot remained at position 2 despite the occurrence of a brief light flash at position 3. (b) Activity from each neuron of the lesioned network and from the light sensor; data for the entire epoch is shown which constituted movements to 6 different sustained lights and inability to move to the brief light flash. (c) FFT data for all five neurons of the lesioned robot over the entire epoch. The value of the corresponding Hurst exponent of each neuron is on the top right.