H. Van Dyke Parunak, James Kindrick, Bruce Irish
Two important components of connectionist models are the connectivity between units and the propagation rule for mapping outputs of units to inputs of units. The biological domains where these models are usually applied are nonconservative, in that a single output signal produced by one unit can become the input to zero, one, or many subsequent units. The connectivity matrices and propagation rules common in these domains reflect this nonconservativism in both learning and performance. CASCADE is a connectionist system for performing material handling in a discrete parts manufacturing environment. We have described elsewhere the architecture and implementation of CASCADE [PARU86a] and its formal correspondence [PARU86c], ] [PARU87a] with the PDP model [RUME86]. The signals that CASCADE passes between units correpond to discrete physical objects, and thus must obey certain conservation laws not observed by conventional neural architectures. This paper briefly reviews the problem domain and the connectionist structure of CASCADE, describes CASCADE’s scheme for maintaining connectivity information and propagating signals, and reports some experiments with the system.