To test tuning methods for the Retina Encoder (RE) of a Retina Implant (RI) as a visual prosthesis for blind subjects with retinal degenerations a suitable simulation of the patient’s evaluative response to RE state alterations must be provided. RE simulates real time retinal information processing and consists of several hundreds of spatio-temporal receptive field (RF) filters to generate electrical signals for ganglion cell (GC) stimulation. We propose a neural network to reconstruct the RE input from a number of consecutive RE output frames. The network can be interpreted as a simulation of a part of the central visual system with GC signals as input and perception visualization as output. We present first results using Evolution Strategies for neural network weight optimization.

Retina Implants for blind subjects with retinal degenerative disorders require retinal information processing in real time with many individually tunable spatiotemporal filters with antagonistic receptive field properties (RF filters). We describe a RF filter structure, function and tuning mechanism as well as the implementation of tunable RF filters in real time on a digital signal processor (DSP) for currently developed retina implants. Each tunable RF filter is capable of implementing a large variety of spatial and temporal mapping operations from spatiotemporal light patterns at the assigned array of photo sensors onto a corresponding single asynchronous stimulation pulse sequence at the output. For this purpose, our tunable RF filters incorporate a wide range of typical RF properties of primate retinal ganglion cells in their functional space. A number of spatial and temporal parameters in the RF filter algorithm is specified in order to assure smooth maneuvering within the functional space as well as tuning of a given RF filter to the desired RF function of a given contacted ganglion cell in a learning phase. Typical spatial and temporal information processing properties of tunable RF filters are presented.

We present a new neural velocity force control scheme for a 6 DOF industrial manipulator ensuring tracking of end effector positions along unconstrained directions and tracking of contact force along the constrained direction to significantly expand the range of manipulator applications. Neural velocity force control is actually feasible even in the case of an extreme stiff environment, which is a quite common situation in industrial applications. A cascaded velocity controller (CVC) ensures the precise approach with a tender impact to the unknown surface. The neural controller is of inverse dynamics type with a force feedforward action and performs an adaptive computation of the inverse manipulator model. Simulation results for a 6 DOF industrial manipulator are reported and show the convenience of the approach for demanding tasks like dismantling or surface tracking inclusive establishing contact to rigid objects with precisely bounded impact forces1.

We developed a novel approach for surface tracking and force/position control based on neural networks. A new concept of neural trajectory optimization NTO will be presented as a part of the neural force/position control NFC and as a very capable and versatile tool for the generation of natural manipulator movements (fast, flexible and smooth). The NTO concept is based on DRBF neural networks, an extension of the RBF type network, to be proposed in this paper. Experimental results of the realtime implementation of NTO and simulation results of its combination with the neural force/position control NFC will be presented. As a testbed we use a 6 DOF industrial manipulator executing demanding tasks such as surface tracking with defined normal force.1

A prestructured neural network, capable of learning and generating smooth multi-dimensional desired trajectories was designed for the control of robotic manipulators. The recurrent network structure combines adaptive parceling of the workspace with a second order differential equation for the relation between force vector and fingertip position. Network performance for storage (using a modified δ-rule) and generation of a given 2-dimensional trajectory was successfully tested. Simulation results showed that additional implementation of adaptive parceling significantly improved the storage accuracy relative to learning of accelerations with fixed parceling.

Neural Networks consisting of localized basis functions are used for approximation of continuous functions. They are known for fast adaptation (Moody et. al., 1989) compared to global networks of multilayer perceptrons. Due to the locality of the Gaussian radial basis functions (GRBF) it is possible to take only a small fraction of all neurons into account without a significant change in network performance. This observation is the reason to search for efficient data structures to find the small fraction of necessary neurons quickly. There are data structures to accelerate the search for the k nearest neighbors of a data point like k-d trees, ball-trees and bump-trees e. g. (Omohundro, 1991). But they all lack the guarantee that all of the k nearest neighbors are found using a single leave of the tree. On the other hand they do not take the finite area of influence of the localized basis functions with the possibility of strongly varying width into account. There exist algorithms for non-radial basis functions, which are better suited for tree based acceleration of recall and learning (Sanger, 1991).

Networks with ’Partition of Unity’ Gaussian radial basis functions (RBF) have achieved considerable attention for their capability to incorporate rule-based knowledge into the network and to extract rule-based knowledge from the network (Tresp et. al., 1993). On the other hand, RBF networks are known for fast adaptation as being used for function approximation. Combining these properties, e. g. prestructuring the network with domain knowledge and refining it with rapid training, is a desirable goal. However, the capability of ’Partition of Unity’ RBF networks to approximate arbitrary functions arbitrarily close on a bounded domain, the essential requirement for universal function approximators, could not be guaranteed so far.

We present a method to use RBF networks for stable adaptive robot control, with the task to track prescribed joint trajectories. The controller architecture is based on the concept of feedback linearization. The RBF networks serve as the model-based controller part. The learning rule is derived from a Lyapunov-based stability criterion to assure global stability at all times. Bounds on network approximation errors, which are essential to validate the stability proof, are established with the help of multidimensional sampling theory. Training the networks with point-to-point movements between randomly generated locations leads to a trajectory independent controller over the complete working range. Simulation results for a planar 4-joint manipulator (4JM) are given to demonstrate the performance of this control method.

Consider a finite n-dimensional space, with an n-dimensional vector S completely describes the “state”, in which a nonlinear system resides; to know the vector of state-changes is a complete representation of the system dynamics provided it is deterministic and of first-order. The totality of all state-changes form a vector field that contains the entire system dynamics. A trajectory in state space is formed by pieceing together infinitesimal steps in the direction indicated by this vector field. The result is a curve whose tangent at each point is always aligned with the vector field (Thompson and Stewart, 1986).

In order to continuously map desired tool center point (TCP) movements onto the corresponding set of joint angles for a 6-DOF robot arm without the need for repeated calculation of the inverse Jacobian we present an adaptive neural network with the following features: Combining these properties, a small, but efficent network architecture evolves, which solves the differential inverse kinematics problem.

We propose a novel scheme, named ‘Bellman Mapping Machine’ (BMM), that aims to extend the scope of reactive robot controllers towards more complex tasks. BMMs are implemented by shallow feed forward networks, that receive as input the compound information about desired action task, present robot state, and short-term predicted cluttered constraints. The street-crossing problem serves as a test-bed for our performance studies: the task there is to generate optimal goal-directed robot motor trajectories, that avoid collisions with moving obstacles, at the same time respecting the robot’s dynamics limitations. We supervise some novel higher order neurocontroller with optimal control examples as computed by Dynamic Programming. We find very efficient representations of the underlying optimal control policy space, as well as sensible generalization to new control situations.

A set of Neural Networks learns matrix components representing mass-coupling, coriolis, and friction forces in an inverse robot model. While obtaining training data from specific random trajectories, the inverse model becomes highly precise over the entire robot working range. Global model precision and subsequently L∞-stability for a controller using the networks in the feedback loop following the learning phase are established by a new method. The approach is demonstrated for a simulated planar 4-joint-machine.

This paper deals with the problem of learning to swing up an inverted pendulum, which belongs to the class of highly nonlinear, non-minimum phase control problems without a general control methodology. It is thus a challenge for reinforcement learning over time (Sutton, 1988).

Concerning the simulation of neural networks with arbitrary topology on distributed memory multiprocessor systems, we introduce our approach to an automatic determination of a near-optimal mapping of neural networks onto a given multiprocessor. Our approach is based on stochastic local search. We propose a decomposition of the mapping into a network partitioning step followed by a placement of partitions onto processors.

Recent high-tech developments in microsystems technology, neural computation, and biosystems technology offer a significant potential for the development, production, and application of a new generation of implantable, learning neuroprosthetic devices. One particularly promising development concerns the international effort towards intelligent visual prostheses. The goal of visual prosthetics is the elicitation of useful visual percepts by means of electrical stimulation although the access to the perceptual domain via the central visual system is still not well understood. For many years, philosophical and psychophysical studies emphasized that visual percepts may be of very different origin. For the purpose of the development of a visual prosthesis, we interpret the function of the central visual system as a sequence of two mapping operations. Our development of a learning retina encoder (RE) evolved in several stages. The retina module RE serves as functional replacement of the retina and is assumed to perform a mapping operation M1 of an input pattern P1 from the physical domain onto the neural domain. In contrast, the adjacent central visual system module is assumed to perform another mapping operation M2 of the M1-output signals in the neural domain onto the perceptual domain thus generation a visual percept P2. During the proposed iterative and perception-based learning process, a human subject (normally-sighted for developments; blind in future applications) provides a perceptual similarity estimate of P2 with respect to P1 as input for a learning algorithm, which in turn modifies the parameter vector of RE until P2 appears sufficiently similar to P1. More specifically, the ensemble of spatio-temporal (ST) filters of RE can be tuned in interaction with a human by means of genetic algorithms (GA). Alternatively, a different retina encoder version, RE*employs a combination of specific ST filter classes as well as a decision-tree algorithm and simulated miniature eye movements. This approach promises significantly improved tuning results. More recent RE development attempt to further enhance the RE properties based on pattern pre-processing of P1, pattern segmentation, selective tuning of temporal pattern presentation, and selective control of clusters of stimulation electrodes. Future commercial success in the field of intelligent, implantable neuroprosthetic devices will significantly depend on the formation of ‘hybrid’ teams of experts from biology, medicine, and technology. Intelligent technology-based therapies, which support or even functionally replace deficient parts of the human nervous system such as specific cases of blindness, deafness, paraplegia, Parkinson disease, and epilepsy are well under way or will become within reach within the next 5 to 10 years.

We studied the conditions for joint information processing of a learning retina implant and central visual system in humans with normal vision in preparation of future retina implant applications in the blind. The visual system was modeled by a retina module (RM) as a learning Retina Encoder (RE) with spatio-temporal (ST) filters and a central visual system module (VM). RE performs a mapping of an optical pattern P1 from the physical- onto a neural domain, whereas VM performs a mapping from the neural- onto the perceptual domain and yields a visual percept P2. Our simulation results suggest that the elicitation of ‘Gestalt’ percepts may be improved by dialog-based RE tuning with evolutionary algorithms and by simulated miniature eye movements. However, considerable efforts in neuroinformatics are still needed to elucidate not only the algorithmic representation of data in the neural domain but also its enigmatic mapping onto the perceptual domain.