Back
to indexJeff Forbes and Timothy Huang and Keiji Kanazawa and Stuart J. Russell
Summary
Create a "belief state" out of a dynamic probabilistic network, then update in real time using temporally invariant networks.
The suitability of this approach toward the urban challenge is huge, but it's a complicated paper to understand. Figures at the end show how the network handles lane changes in several situations. This work is also quite old, so enhancements have probably been completed.
Methods
Belief state is the current state upon which a decision for the next step will be made. Belief states are nodes in a dynamic probablistic network, which is an acyclic graph with arcs drawn between nodes indicating "casual connections amoung variables". Nodes include tables of conditional or prior probabilities. At each time slice, add a layer onto the network, which becomes the leading edge. Structure provides ability to track target vehicle's next position like a Kalman filter, with the enhancement that the vehicle's path may be any direction or more widely varied (i.e. avoid debris). Seperate networks are maintained for each vehicle.
Decision making is performed by a POMDP - partially observable Markov decision process. The optimal decision is a function of the current belief state, using the joint distibution over all possible actual states. A utility function is applied over all modes to find the maximum. The utility function can include reward nodes, which incorporate progress toward goal or negative progress toward goal. Node connections are paritally based on hand-constructed trees of rules.
Keywords
dynamic probabilistic network, temporally invariant networks
Rating
8
Bibtex Entry
@inproceedings{ forbes95batmobile,
author = "Jeff Forbes and Timothy Huang and Keiji Kanazawa and Stuart J. Russell",
title = "The {BATmobile}: Towards a Bayesian Automated Taxi",
booktitle = "{IJCAI}",
pages = "1878-1885",
year = "1995",
}