Knowledge Compilation for Online Decision-Making: Application to the Control of Autonomous Systems

The objective of the thesis was to study how knowledge compilation could be applied to the control of autonomous systems. We focused on realistic planning problems, which often involve variables with continuous domains or large enumerated domains (such as time or memory space). We oriented our work towards the search for target compilation languages expressive enough to represent such problems.

In a first part of the thesis, we present various aspects of knowledge compilation, as well as a state of the art of the application of compilation to planning. In a second part, we extend the BDD framework to real and enumerated variables, defining the interval automata (IAs) target language. We draw the compilation map of IAs and of some restrictions of IAs, that is, their succinctness properties and their efficiency with respect to elementary operations. We describe methods for compiling into IAs problems that are represented as continuous constraint networks. In a third part, we define the target language of set-labeled diagrams (SDs), another generalization of BDDs allowing the representation of discretized IAs. We draw the compilation map of SDs and of some restrictions of SDs, and describe a method for compiling into SDs problems expressed as discrete continuous networks. We experimentally show that using IAs and SDs for controlling autonomous systems is promising.