SeqNet: a connectionist network for rule driven sequential problem solving

Summary form only given, as follows. A method for enhancing connectionist networks with sequential processing abilities is proposed. Sequentiality is achieved without global operations applied to the network, and is considered nontrivial since it can arise without dependence on changes to the network´s input by the environment. The method is realized physically in the architecture of SeqNet, a connectionist model which generates sequential steps to solve problems in a simple blocks-world manipulation domain. SeqNet uses explicit microrules in the form of condition-action pairs loosely similar to those in a symbolic production system, but the rules are implemented using the same type of simple computing unit used for internal representations of the environment. A local attractor energy minimization procedure drives the microrule matching process, while the nontrivial sequentiality arises from the controlled absence of weight symmetry between units. The result is a connectionist architecture which inherits some of the strengths of both traditional symbolic approaches to sequential problem solving and connectionist constraint satisfaction approaches.<>