 |
Center for Nonlinear Studies |
Dynamic programming is a broadly applicable methodology for sequential decision making, but suffers from exponential growth of computational requirements as the problem size increases. This has led to extensive work on approximations over the last twenty years. One key idea is to use an (approximate) scoring function to select decisions in complex dynamic systems, arising in a broad variety of applications from engineering design, operations research, resource allocation, finance, etc. This is much like what is done in computer chess and computer backgammon, where positions are evaluated by means of a scoring function and the move that leads to the position with the best score is chosen. Neuro-dynamic programming/reinforcement learning provides a class of systematic methods for computing appropriate scoring functions using neural network-like approximation schemes and simulation/evaluation of the system's performance. Another important idea is to use heuristics to compute on-line the values of an approximate scoring function, and is well-suited for large discrete optimization problems. The talk will overview these methodologies and discuss recent work.
Host: Frank Alexander