ICAPS05 WS6 - icaps 2005

icaps05.uni.ulm.de

ICAPS05 WS6 - icaps 2005

tion. On the originating side of the macro action, desire

states require recalculation. On the resulting side of the

macro action, previously calculated values can be reused.

Controlling Uncertainty through Coordination

While the uncertainty of the UAV Surveillance domain

is captured by decision-theoretic planning, the uncertainty

from the action of other agents is not. Using the methods

of goal addition, removal, and modification, coordination

among the agents can be used to reduce uncertainty. Based

on the locations of the UAVs and any messages passed between

the agents, an agent can adjust the expected reward

values in its desire space to reflect the probable actions of

other agents.

Different levels of coordination may be employed, depending

on the requirements on resource consumption and

solution quality. The purpose of coordination is to reduce

uncertainty about the expected rewards for visiting each target.

Four types of coordination are examined and their ability

to reduce uncertainty evaluated: (1) no coordination,

(2) location-based inference, (3) communicated inference,

and (4) explicit partitioning.

With no coordination, the agents operate without any

knowledge of the other agents in the system. This option

requires no additional computational resources or communication

on behalf of the agents. Since the agents have no

awareness of the other agents, they tend to operate redundantly,

often attempting to visit the same target. This situation

reflects the most uncertainty.

Location-based inference and communicated inference

both produce an implicit partitioning of the goals, reducing

the overlap in work performed by the agents when compared

to no coordination. Location-based inference uses only information

about the physical location of the UAVs and the

targets. Targets that are closer to other agents have their

expected rewards reduced due to the increased probability

that the other agents will visit those targets first. Communicated

inference is similar to location-based inference, but the

agents calculate which are their preferred targets and communicate

those preferences to the other agents. The benefit

of this over location-based inference is that the agents

can take their paths (i.e., their future locations) into account

when calculating their preferences instead of just their

present location. In these two situations, the agents suffer

somewhat less uncertainty than in the case of no coordination.

With explicit partitioning, the agents negotiate an allocation

of the goals to respective agents effectively reducing the

overlap to zero. One drawback of using explicit partitioning

is an increase in both communications the additional computational

resources needed to calculate and negotiate the

partition. Also, this method can result in commitments far

into the future, reducing performance of the agents restricting

the ability to adapt to changing conditions quickly. This

situation represents the least uncertainty.

Figures 6 and 7 compare the four coordination mechanisms

described above. In each case, three agents are used

to cover a battlefield. Targets are added to random locations

Target Hit Rate

1

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0

No

Coordination

Comparison of Solution Quality

Location

Inference

Communicated

Inference

Coordination Mechanism

Explicit

Partitioning

Easy

Hard

Figure 6: Comparison of the quality of solution as a percentage

of the rewards received by the multi-agent system

Time to Target

160

140

120

100

80

60

40

20

0

No

Coordination

Comparison of Efficiency

Location

Inference

Communicated

Inference

Coordination Mechanism

Explicit

Partitioning

ICAPS 2005

Easy

Hard

Figure 7: Comparison of the efficiency of solution as an average

of the costs incurred per goal

on the battlefield at regular intervals. Difficulty of coverage

was set for the agents by the speed at which targets are

added. Targets have a given lifetime after which, if they

have not been visited by an UAV, they are removed by the

mission commander. If this occurs, it is counted as a missed

target. Figure 6 shows the effect of the coordination mechanisms

on the ability for the agents to spread out across the

battlefield. The results show that explicit partitioning is the

best, while the implicitly partitioning of location inference

and communicated inference are slightly better than no coordination.

Figure 7 shows the efficiency of the agents at

retrieving their rewards, measuring the distance travelled on

average to visit each target since cost is dependent upon distance.

Increasing the amount of coordination reduces the

distance travelled, meaning there was less overlap in the actions

of the agents due to less uncertainty about the actions

of other agents.

Workshop on Planning under Uncertainty for Autonomous Systems 55

More magazines by this user
Similar magazines