Multi-objective Optimization and Mixed-Horizon Decision-Making for Autonomous Vehicles
Abstract
Project Summary The interval programming (IvP) model is a method for representing and solving multi-objective optimization problems and the IvP Helm is an autonomous decision making application that uses the IvP model to apply multi-objective optimization for unmanned vehicle autonomy. The work proposed here involves advances in the fundamental representation scheme of the IvP model and solution algorithms, with the objective of providing advances in the types of decision making available to unmanned vehicle autonomy solutions. There are three proposed advances to the IvP model and solution algorithm. (a) Constraint reasoning. Presently an IvP problem instance is comprised of an unconstrained combination of weighted utility functions where effectively constraining out intolerable decisions is achieved by ensuring sufficiently dominant weights applied to component functions. The proposed extension to the model will allow explicit identification of infeasible solutions that is not dependent on weights. Section 1.4 (b) Non-additive weight combinations. Presently the solution to an IvP problem is defined by a weighted combination of objective functions. In certain cases, a pair or group of functions may be more aptly combined in a non-additive, e.g. max-value manner. Section 1.5. (c) Solution threshold grouping. The IvP solution algorithm presently finds the single best next solution. For the reasons discussed in Section 1.9, it may be useful to also identify the group of solutions exceeding a certain threshold, as a means for seeding multi-step path planning algorithms. There are three proposed advances to the IvP Helm autonomy architecture. (a) Approximate vehicle dynamics. This involves a new method for approximating and representing vehicle turn characteristics that does not required hydrodynamic modeling. Trajectory estimation will be built into the geometry and autonomy libraries of the IvP helm to better evaluate candidate maneuvers. Section 1.6. (b) Automatic discovery of approximate vehicle dynamics. Autonomy missions are investigated to minimize the time and operation area needed for a vehicle to sufficiently experiment various turn-speed combinations for a full characterization based on both actually observed and interpolated components. Section 1.8. (c) Mixed-horizon decision making. The IvP model, as presently used in an autonomous helm, provides decisions for the next time step without explicit consideration of following decisions further into the future. In this thread we will explore using IvP for seeding the initial sets of decisions for multi-stage path planners such as RRT or D* planners.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Aug 12, 2016
- Source ID
- N000141512227
Entities
People
- Michael Benjamin
Organizations
- Massachusetts Institute of Technology
- Office of Naval Research
- United States Navy