Population based optimization for variable operating points

Finding optimal inputs for a multiple input, single output system is taxing for an system operator. This work presents a population-based optimization to create sets of functions to approximate a locally optimal input as an operator selects an output. Output and cost functions are modeled by neural networks. Neural network gradients are used to optimize a population of agents by minimizing the cost for the agent´ s current output. When an agent reaches an optimal input for its current output, additional agents are generated to step in the output gradient directions. The agent then settles to the local optimum for the new output value. The set of associated optimal points forms a inverse function, via spline interpolation, from a desired output to an optimal input. In this manner, a locally optimal function is created for each settled agent. These functions are naturally clustered in input and output spaces allowing for a continuous optimal function. The best cluster over the anticipated range of desired outputs can be chosen and the process optimized on-the-fly to respond to different set points. Results are shown for a diverse set of functions.