Decision making under uncertain conditions

Prof. Vadim Indelman | Aerospace Engineering


Automation, Mobility and Aerospace | Information and Computer Science

The Technology

A system and methods are provided for decision making under uncertainty, for selecting an optimal action from among multiple candidate actions in belief space planning (BSP). BSP is a scalable approach for planning under uncertainty. Today, the two processes, inference and decision making, are being treated separately: the inference stage maintains a belief over variables of interest given available information, while decision making under uncertainty is entrusted with determining the best next action(s) to realize a certain objective. Since the system is unobservable and hence inference about the probability distribution over the state space, also called belief, is performed based on the future observations.

Most of the state-of-the-art approaches assume either that the position is known or that the perceptually-aliased objects are sufficiently distinguishable. Additionally, current decision-making problems are based on the same fundamental paradigm. At each planning session, the precursory inference is being propagated with different hypotheses over information that might be acquired in succeeding inference stages. Based on the sophistication level of the predictor that created these hypotheses, the resemblance to the actual information can range from completely different to nearly identical.

This decision-making approach as it may, has to solve numerous inference problems in order to determine the optimal actions. The novelty lies in considering the uncertain data association within the belief space planning. This results in a mathematically elegant way to incorporate data associations and also compute the weights of the resulting hypotheses.


  • Autonomous systems in an unknown environment with several look-alike features
  • Real-time operation in high dimensional state spaces
  • Study properties of geometrical objects invariant to certain deformations
  • Principled way to account for uncertain data association as opposed to any domain-specific ad-hoc approach
  • No additional computational costs
  • Updating inference (belief) with incoming data

Applications and Opportunities

  • Computationally efficient approaches for decision making under uncertainty are useful for a range of artificial intelligence applications, among others guidance of autonomous systems (navigations, autonomous car industry, autonomous robotics, quadrotors, defense industries, etc.)
  • Ensure correct mapping and localization
arrow Business Development Contacts
Motti Koren
Director of Business Development, Life Sciences