This problem in artificial intelligence is also called "belief space planning". Solving this problem, i.e. calculating the set of optimal actions to achieve the goal, requires an estimate of the potential actions under some reward or price function such as distance to the goal or a measure of uncertainty
In the age of autonomous cars, industrial robots and intelligent systems that assist humans in various situations, time and computing resources are valuable factors. These systems are required to respond quickly to circumstances in a changing environment and in a state of missing information (uncertainty). In addition, economic constraints limit the complexity of elements such as hardware, and systems must be cheap enough for potential consumers to pay for them. Research conducted at the Faculty of Aeronautics and Space Engineering at the Technion and published in the International Journal of Robotics Research presents a theoretical and computational breakthrough in this context: the simplification of planning and decision-making problems under uncertainty in a way that reduces the amount of data that the computer is required to analyze.
The research was led Prof. Vadim Indelman Head of the Laboratory for Autonomous Navigation and World Sensing (ANPL) in the Faculty of Aeronautics and Space Engineering, and Han Elimelech, who recently completed a doctorate in the technical program for autonomous systems (TASP). "We show that we can significantly reduce the required computing power without harming the success of the task," the two explain. "Furthermore, we show that it is possible to reduce the computational effort even further for a certain possible performance penalty - a penalty that our approach can estimate online. In the era of autonomous cars and other robots, this is an approach that may enable online autonomous decision-making in challenging scenarios, reduce response times and save significant costs in hardware and other resources."
Prof. Indelman's research deals with autonomous decision-making under conditions of uncertainty - a fundamental problem in artificial intelligence and robotics. This ability is especially essential for autonomous agents that are required to operate autonomously and reliably over time, under conditions of uncertainty and in a changing environment. Furthermore, in many cases the agent does not have direct access to the state variables of the problem and operates based on a probability distribution or Faith (belief). This distribution expresses the knowledge the agent has about himself and his environment based on probabilistic models, actions performed and measurements received from sensors in his possession.
One of the main research directions in Prof. Indelman's research group is online and reliable decision-making under these conditions. This problem is also called "belief space planning". Solving this problem, that is, calculating the set of optimal actions to achieve the goal, requires an estimate of the potential actions under some reward or price function such as distance to the goal or a measure of uncertainty. It is worth noting that this challenge requires predicting how Faith will evolve in the future for different possible actions while predicting different scenarios of the future. Therefore, making decisions under these conditions is computationally expensive, which challenges the autonomous operation of smart agents in real time. Furthermore, in problems where there are many state variables (for example, when the environment changes or is unknown in advance) the computational challenge is even more acute. As mentioned, to all of these are added the economic consideration and time and calculation capacity constraints, which require a reduction of the required computing resources, and therefore simplification of planning problems under uncertainty is an important goal in these research directions.
Prof. Indelman's research group addresses all these aspects in the development of simplification approaches that allow solving the above problems in a more computationally efficient way. This is, for example, through thinning (sparsification) of matrices. An essential aspect of this type of approach is receiving performance guarantees - that is, whether the performance can be damaged, and to what extent, as a result of the abstraction process. Indeed, in the current article, the researchers lay the foundations for solving decision problems through abstraction and, moreover, show that these approaches can lead to significant savings in calculation times, and this without significantly impairing the fulfillment of the task.
Doctoral student Chen Elimelech, who led the research, recently won an "award for outstanding research work for a PhD" on behalf of the Israeli Smart Transportation Research Center (ISTRC).
The research was supported by the National Science Foundation (ISF).
for the article in the journal International Journal of Robotics Research
