• Tidak ada hasil yang ditemukan

Communication Control

Dalam dokumen Autonomous Agents and Multi-Agent Systems (Halaman 33-36)

Information-Based Control of Decentralised Sensor Networks

4. Examples and Applications

4.3. Communication Control

This example considers active information flow in decentralised sensor networks.

This refers to smart dissemination of information over a network, with due regard to the local utilities of receiver/sender nodes or the global utility of the system, as well as communication resource constraints. It may be possible to organise such flows on the basis of simple rules or heuristics, but these often require careful

‘hand-tuning’ and are unlikely to offer much flexibility.

A more general approach, developed under the SEAS DTC project AA009 (see Section 1.1 for details), is highlighted in this section. A typical motivating problem is shown in Fig. 5. This is a surveillance mission scenario, involving mul- tiple decentralised and heterogenous agents and users, operating in an uncertain and hostile environment, where communications are limited. The problem is what information should the agents and users exchange to ensure timely execution of their mission goal(s)?

There are two variants of IBC that can be applied to this problem. Both are underpinned by Bayesian Decision Theory but differ in their implementation de- tails depending on whether the sensor nodes have a common objective or unequal

Figure 5. Surveillance scenario involving multiple sensing agents and multiple users. Of interest is how the agents and users should interact to gather relevant data in a rapid and effective manner.

objectives. Each variant was evaluated in simulation against a target identification problem (represented by a Bayesian Network). In each case there was an empha- sis on providing scalable decentralised solutions. This ruled out the conceptually simple approach of replicating the centralised solution at each node, because that requires each sensor node to acquire all the other sensor nodes expected utilities.

The first method, developed for common local utility functions, is charac- terised by informationpush in reaction to what is known by a transmitting node about a receiving node’s information requirements [6]. Two smart steps are re- quired to implement this method:

1. The probabilistic world model is represented in a compact factored form known as a junction tree. This enables an algorithm for inference within

Bayesian Networks to also perform inference between Bayesian Networks at separate locations.

2. An efficient communication protocol is used to minimise inconsistencies be- tween the probabilistic estimates of state maintained by each node. Specifi- cally, Kullback-Leibler divergence is used to monitor and prioritise informa- tion flow in the system.

The main strength of this method is its efficiency: it uses fewer resources than competing methods when resources are unconstrained, and provides faster conver- gence and increased accuracy when communication is constrained. However, the intrinsic weakness of this method is its assumption of a common objective. In mil- itary operations, it is quite likely that sensor nodes will have different objectives which are related to their local context. Moreover, each node’s local objective is unlikely to be known to the other nodes.

The second method, developed for different objective functions, is charac- terised by informationpull, which is a proactive advertisement by the sensor nodes for information that supports their local objectives [7]. As in the information push scheme, two smart steps are required to implement this method:

1. A means of generating advertisements that can be interpreted by receiving agents and used as a basis for information gathering and communication decisions. A suitable advertisement is a vector containing the average utility change for every possible request, normalised to form a set of priorities over actions.

2. An efficient “in-network” scheme for aggregating advertisements and informa- tion. By using a tree communication topology and storing the advertisements received on each link, it is possible to formulate and propagate an aggregated advertisement. In this way agents can use “link demands” to steer informa- tion toward the desired destinations.

In common with the information push method this method is also scalable because it communicates and fuses estimates rather than sensor data, it exploits structure in the world model, and it prevents stale data from being re-transmitted. In ex- change for its scalability and flexibility with respect to multiple objectives, the information pull method trades optimality in performance.

These concepts for communication control were exposed to a simple experi- ment in which five (simulated) airborne sensor nodes were tasked with maintaining surveillance of targets in their own areas of interest and reflecting this self-interest in their private utility functions. The sensor nodes are ignorant about each other’s utilities. Within the scenario there are three sensor types. Each sensor can distin- guish different target attributes, but evidence from all three sensors is required to positively identify a specific target type. The performance of each sensor is captured using a naive Bayes classifier.

The experiment compares performance for the information push/pull ap- proaches in terms of the Shannon information utility averaged over nodes. The

Figure 6. The mean Shannon information utility for a bandwidth-limited sensor network implementing information push and pull algorithms as shown. Information pull performs better (i.e., has lower Shannon information) due to the differing objectives of the sensor agents

results are shown in Fig. 6. The information push approach assumes common ob- jectives and in this scenario, where the sensors have their own private interests that are unknown to other sensors, it is outperformed by the more flexible information pull approach.

Dalam dokumen Autonomous Agents and Multi-Agent Systems (Halaman 33-36)