Sequence generation with connectionist state machines

Summary form only given, as follows. Backpropagation networks with state memories were trained to generate sequences of discrete events. In one study, sequential networks were trained to produce 'verbal' descriptions of objects in a microworld. In a second set of studies networks were trained to manipulate a blocks world. One version required the network to generate a sequence of actions for manipulating the blocks in response to instructions. A second version trained networks to generate actions to move blocks from an initial configuration to a goal state. In a final set of studies, networks generated strings of features. These networks were shown to take advantage of the structure of the output sequences and to apply output rules when generating sequences.<>