Room 206 (2nd floor, badged access)
13 June 2019 - 14h00
Using neural networks to improve performances of correct-by-construction controllers
by Simon IOSTI from Verimag
Abstract: The use of neural networks is usually considered as antinomic with correct-by-construction paradigms. This situation can be thought of as illustrating the trade-off between performance and accuracy. Recent researches and ideas have nevertheless been put forth to investigate mixed approaches allowing for the use of deep learning methods in correct-by-construction controller synthesis (see, for example, the ETLAS project recently presented by Saddek Bensalem).
This work is a first step towards such mixed methods. Consider the following distributed control problem: several processes running in parallel have a common goal to enforce, and are allowed to synchronize intermittently to take decisions towards this goal. The drawback is that synchronizations are costly, and may result in insufficient information to be able to take a decision; hence synchronizations should be minimized, and their efficiency maximized, in order to improve the performances of a distributed controller. Knowledge-based algorithms are known to build such controllers when they exist, but the question remains to improve their performance regarding synchronizations. We propose an approach using recurrent neural networks (RNNs) to optimize these controllers.
In this talk, I will present formally a simpler model of a system that must control an unknown environment. To do so, the system has the ability to attempt actions, and the environment must comply if this action is available to it. The only information available to the system is the success or failure of such control attempts, the state of the environment staying invisible. The system is given an RNN whose purpose is to learn the structure of the environment by probing it with control attempts, aiming at reducing the overall number of failed such attempts. I will present RNNs and the motivations for using those in this setting, several examples showing various difficulties to the learning that might be encountered, and present our experimental results.