Adding Neural Network Controllers to Behavior Trees without Destroying Performance Guarantees
share
In this paper, we show how controllers created using data driven designs, such as neural networks, can be used together with model based controllers in a way that combines the performance guarantees of the model based controllers with the efficiency of the data driven controllers. The considered performance guarantees include both safety, in terms of avoiding designated unsafe parts of the state space, and convergence, in terms of reaching a given beneficial part of the state space. Using the framework Behavior Trees, we are able to show how this can be done on the top level, concerning just two controllers, as described above, but also note that the same approach can be used in arbitrary sub-trees. The price for introducing the new controller is that the upper bound on the time needed to reach the desired part of the state space increases. The approach is illustrated with an inverted pendulum example.
READ FULL TEXT