A connectionist approach to direct dynamic programming control

Direct dynamic programming controllers can be constructed with artificial neurons. Doing so facilitates the storage of content-addressable dynamic programming cost values only in regions of the state space that are actually passed through during control of the plant, thereby reducing total memory usage. Further, by connecting the neurons in different ways, neighboring states can be amalgamated together and treated uniformly. This gain reduces the overall memory requirement of the direct dynamic programming controller and can also reduce its learning time by facilitating generalization. A number of experiments with such neural networks designs are described, along with a particularly promising design that has emerged from this research. This new approach, transition point dynamic programming, operates by placing neurons only at states where a change in the control signals is required