No label defined (Q4250893)

There has been a big dispute about the effectivity of neural networks applications in solving real problems in comparison to standard techniques. The reviewed paper is a fine example illustrating how the intelligent exploitation of neural networks can beat the standard techniques in the case of adaptive stochastic path planning for robots. Mathematically, the problem of trajectory planning for robots under stochastic uncertainty can be solved by the methods of stochastic optimization. In order to improve the control of the robot, the velocity profile and the geometric path must be adjusted. Consequently, one has to solve an adaptive stochastic trajectory planning problem. The authors present how to solve such an optimization problem by means of the finite element methods. This seems not appropriate to be used in real-time computation, as industrial robots move very rapidly. In fact, to compute a given geometric path for the point-to-point path planning should take a rather long time, preventing to use such methods on-line. But such a real-time optimization problem can be solved intelligently by means of neural networks. Here neural networks are used for off-line solving of appropriate problems by learning adequate parameters to be used later. If the neural network is trained off-line with known data, for the novel, actual data it will yield on-line very fast optimal values of required parameters. In this sense neural networks already trained, work very fastly and the whole method can be used on-line in real-time. The numerical results presented by the authors illustrate this strategy remarkably. The paper is written very nicely, in a transparent way. For the entire collection see [Zbl 0912.00035].