In this work it is proposed a self-organizing network, called Competitive and Temporal Hebbian (CTH) network, capable of learning and recalling complex temporal sequences. The CTH network handles sequences in which an item occurs many times or is shared with other sequences. In both cases, uncertainties occur during recall, but context information are used to resolve them. Competitive synaptic weights encode the static portion of a sequence, while the temporal order is encoded by lateral weights. The CTH network saves memory space since only a single copy of each repeated/shared sequence item is stored. Furthermore, a redundancy mechanism improves the robustness of the network against noise and faults. A distributed control platform was used to evaluate the CTH network in trajectory planning for real time, point-to-point control of trajectories of a PUMA 560 robot. The proposed system is compared with other neural network based approaches.
Unsupervised neural networks; temporal sequences; robotics; trajectory planning; distributed control
