8.7 DISCUSSION
This section discussed numerous coordination strategies and presented some results indicating that new coordination techniques based on solving MDPs can exceed the level of safety provided by the current TCAS logic with less disruption to the pilots. This section demonstrated that, in the presence of realistic sensor noise, the Joint-Cent and Uncoord-Compat strategies are viable methods of achieving coordination. Future work will examine the use of incorporating communication into the model. This section only considered evaluation of the joint MDP using the synchronous action set. Future work could reveal the extent to which strategies that use the solution of the joint MDP can be improved through the use of di.erent joint action spaces.
The Uncoord-Compat strategy may be the most appealing way of providing coordination because it can preserve the current communication mechanism of TCAS. The future system would transmit VRC messages potentially inhibiting the advisories from which the intruder could select. An intruder equipped with the same system or with any prior version of the TCAS logic would be able to process the coordination message. Though the Uncoord-Compat strategy does not represent the optimal solution to the problem of coordination, it has been shown to perform remarkably well in realistic encounter situations. It does not require o.ine planning over the joint MDP; no additional cost table beyond that for the uncoordinated MDP would have to be loaded on board the aircraft.
9. MULTIPLE THREAT ENCOUNTERS
The previous sections of this report have focused entirely on resolving encounters with single in-truders. Although encounters involving two or more intruders are relatively rare in the current airspace, the collision avoidance logic must be robust to such situations. This section generalizes the methods introduced earlier in this report to multithreat encounters.
Ideally, a solution would be obtained from an MDP modeling an arbitrary number of aircraft. Unfortunately, each additional intruder would require introducing at least two variables to capture their relative altitude and vertical rate. Intruders with collision avoidance systems would require additional
advisory
state
variables.
Augmenting
the
MDP
in
Section
3
would
require
at
least
two
orders of magnitude more processing time to compute the optimal cost table. Storage of the table would require at least two orders of magnitude more memory, which is not currently feasible.
One strategy for handling large MDPs is to decompose the decision making to multiple agents
[50].
In
the
collision
avoidance
problem,
each
agent
may
be
assigned
a
particular
intruder
to
avoid. They independently make recommendations whether to alert and which advisory to issue, and some strategy is employed to make the best overall decision. Several di.erent kinds of strategies have been proposed, including command arbitration and utility fusion. TCAS uses a method that resembles command arbitration, but the MDP approach permits utility fusion, which may be more appropriate for resolving multithreat encounters.
9.1 COMMAND ARBITRATION
In command arbitration, an arbiter selects one of the actions recommended by one of the agents. One of the simplest command arbitration strategies is to accept the recommendation of the agent whose intruder is closest. In multithreat encounters, one intruder is usually much closer than the others, and so resolving NMACs sequentially in this manner may be satisfactory most of the time. However, it is easy to construct situations where this strategy fails.
Figure
32
shows
the
action
regions
for
the
closest
command
arbitration
strategy
中国航空网 www.aero.cn
航空翻译 www.aviation.cn
本文链接地址：Robust Airborne Collision Avoidance through Dynamic Programm(44)
