Abstract
Many schemes have been put forward for general-purpose learning machines capable of performing a variety of tasks, but little work has been done on direct comparison of the performance of these with the most adaptable of all controllers, the human operator. A suitable environment for such a comparison is described here, and experiments with several different adaptive controllers are reported. The most successful of the automatic adaptive controllers consists of a series of stochastic estimators, the outputs from which determine the probability that the controller will take a certain action for each region of the environment's statespace. The results show that human operators can achieve good control by estimating derivative information from a continuous display of the state of the environment, but are badly affected by random external disturbances. Even the best automatic controller considered can use only coarsely quantized input information, and so has poorer control, but it is relatively immune to external disturbances.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have