Adaptive learning algorithm

Judson and Rabitz [60] have provided a numerical demonstration of an existence theorem for feedback control in the guiding of the evolution of the state of a system. The example they consider is the transfer of 100% of the population from the vibrationless ground rotational state of KC1 to the vibrationless state with j = 3, m = 0, by a suitable field. The novel idea they exploit is to use the population transfer generated by a trial field as input to an adaptive learning algorithm for comparison with the desired popula-... 

All of this can be controlled using computer software, which should also include an adaptive learning algorithm, capable of computing the next step towards producing the optimum product yield and then tailoring the laser pulse in an appropriate way (this is termed adaptive closed-loop control). After each laser pulse,... 

Riedmiller, M., Braun, H. RPROP-a Fast Adaptive Learning Algorithm. In Proceedings of ISCIS VII, Universitat (1992)... 

Carpenter, G. A., Grossberg, S., and Rosen, D. B., ART2-A An adaptive resonance algorithm for rapid category learning and recognition, Neural Network 4, 493 (1990). 

Genetic algorithm A learning algorithm used to maximize the adaptability of a system to its environment. The method, based on the genetic processes of re-... 

Neural network learning algorithms BP = Back-Propagation Delta = Delta Rule QP = Quick-Propagation RP = Rprop ART = Adaptive Resonance Theory, CP = Counter-Propagation. 

Figure 19.1 Diagram showing the arrangement for closed-loop learning control. Following a femtosecond laser pulse, the products of the photochemical process are detected and compared with the user-defined objectives stored on the computer. A learning algorithm then calculates the modified electric fields required to shape the laser pulse and further optimize the yield of the desired product. Cycling through the loop many times gives the optimum pulse shape and best product yield. Adapted from Brixner et o/, Chem. Phys. Chem., 2003, 4 418, with permission of John Wiley Sons Ltd...

It is important to emphasize that, in the above examples, knowledge of the PES was not required for the optimization process. The adaptive-control learning algorithm explores the available phase space and optimizes the evolution of the wave packet on the excited state PES without any prior knowledge of the surface. Thus, the intrinsic information about the excited-state dynamics of these polyatomic systems remains concealed in the detailed shape and phase of the optimized pulse. Inevitably, however, scientific curiosity, together with a desire to imder-stand how chemical reactions can be controlled, has led to pioneering studies that aim to identify the underlying rules and rationale that lead to a particular pulse shape or phase relationship that produces the optimum yield. 

Luo, Z. (1991). On the convergence of the LMS algorithm with adaptive learning rate for linear feedforward neural networks, neural computation. 

Learning algorithms for neural networks are usually methods that minimize the mean square of system error iteratively by adapting the weights. 

The component techniques of soft computing are not competitive, but complementary. Much research has been done to study the ways this complementarity can be exploited. Each of the components has features to offer a potential partnership. Systems that have such a partnership are called hybrid systems . Fuzzy logic uses the concept of computing with words, it deals with imprecision and information granularity and is an important tool for approximate reasoning. Neural networks learn and adapt. Genetic algorithms make use of a systemized random search and are an important tool for optimization. These three may be combined in different ways, as described below. 

---