Документ взят из кэша поисковой машины. Адрес оригинального документа : http://www.stsci.edu/institute/itsd/information/streaming/archive/IWPSS2006/AndrewCarrel102406Hi_supporting/AndrewCarrel102406.pdf Дата изменения: Tue Sep 25 16:26:37 2012 Дата индексирования: Wed Feb 6 05:02:27 2013 Кодировка: Поисковые слова: m 35

Adaptive Resource Modelling for Autonomous Planning & Scheduling
Andrew Carrel, Phil Palmer Surrey Space Centre, University of Surrey, UK

5th International Workshop on Planning and Scheduling for Space Baltimore, 24th October 2006


Contents Resources types in operations planning. The NEAT planner/scheduler. Example applications of NEAT. Basic adaptation for planning. Neural network resource modelling. Test results. Conclusions & future work plans.
Adaptive Resource Modelling 2 IW PSS, October 2006


Planning with Resources User goals achieved by operations. Operations are groups of activities. Activities have temporal constraints. Activities use resources:
Non-depletable resources Depletable resources

Adaptive Resource Modelling

3

IW PSS, October 2006


NEAT Planning/Scheduling NEAT (Near-optimal Evolutionary Autonomous Task-manager) designed for onboard autonomy.
Evolving Population fittest member Schedule Construction

Valid activity schedule: commands for next cycle
Adaptive Resource Modelling 4

discard
IW PSS, October 2006


Evolutionary Optimisation
Chromosomes are operation permutations. Schedules are constructed using backtracking. Only valid schedules are generated. Evolutionary search never terminates.
Adaptive Resource Modelling 5

Population of Chromosomes Schedule Construction Fitness Function Parent Selection Crossover & Mutation

IW PSS, October 2006


Test Scenarios Mars Rover
Mobile platform visiting science targets Battery charge & memory are depletable resources

UK-DMC Imaging Satellite
LEO, Sun-synchronous orbit Battery charge & 2 в memory are depletable resources
Adaptive Resource Modelling 6 IW PSS, October 2006


Rover Scenario

Adaptive Resource Modelling

7

IW PSS, October 2006


Rover Scenario

Adaptive Resource Modelling

8

IW PSS, October 2006


UK-DMC Imaging Scenario

94 targets taken from imaging log for September 2004. NEAT manages resources effectively. NEAT shows 31% improvement in science return.
120 100 80 60 40 20 0 Existing P l an n i n g NEAT

Adaptive Resource Modelling

9

IW PSS, October 2006


Dependency on Models Planning ahead requires estimating future resource levels. Projection of resources needs:
1. 2.

Values for current resource levels. Models that estimate the effect of future activities on resources levels.

Adaptive Resource Modelling

10

IW PSS, October 2006


When the model is wrong? Two examples tested on the UK-DMC scenario:
1.

2.

Damaged solar array producing less power than expected. Damaged antenna downlinks data at a much lower rate than expected.

Adaptive Resource Modelling

11

IW PSS, October 2006


Damaged Solar Array Open Loop:
100 95

Basic Feedback:
100 95

Battery Charge (%)
0 1 2 3 4 5 6 7 8 9 10

Battery Charge (%)

90

90

85

85

80

80

75

75

70

70

0

1

2

3

4

5

6

7

8

9

10

Time (days)

Time (days)

damage occurs
Adaptive Resource Modelling 12

damage occurs
IW PSS, October 2006


Damaged Antenna Open Loop:
Failed Images
6 4 2 0

0

1

2

3

4

5

6

7

8

9

10

Basic Feedback:
Failed Images
6 4 2 0

Time (days)

0

1

2

3

4

5

6

7

8

9

10

Time (days)

Adaptive Resource Modelling

13

IW PSS, October 2006


Neural Network Learning

Evolving Population

Schedule Construction
Model corrections commands

resource Spacecraft data

Neural Network
Adaptive Resource Modelling 14 IW PSS, October 2006


Neural Network Learning Inputs: Model estimates
Depletable resource rates Non-depletable resources
a x
1

u

1

a

2

u2

Outputs: Model corrections
Depletable resource rates Non-depletable resources

inputs W
1

outputs W hidden layer
2

Training method: Back-propagation.
Adaptive Resource Modelling 15 IW PSS, October 2006


Damaged Solar Array

Battery Charge (%)

Neural network a d a p ts r e s o u r c e model over period of <1.5 days. Resource management fully recovered.

100

95

90

85

80

75

70

0

1

2

3

4

5

6

7

8

9

10

T ime (days)

Adaptive Resource Modelling

16

IW PSS, October 2006


Damaged Antenna

Failed image capture activities completely eliminated. Neural network successfully models reduction in downlink c a p a c ity .

Adaptive Resource Modelling

17

IW PSS, October 2006


Conclusions Autonomous planning/scheduling is very dependent on resource models. A neural network can be used effectively to capture consistent deviations in resource behaviour. NEAT can use this to adapt future plans as resource behaviour changes.

Adaptive Resource Modelling

18

IW PSS, October 2006


Future Work Test this method on the rover scenario (e.g. damaged wheel/motor). Investigate alternative neural network inputs (e.g. subsystem on/off, modes). Assess tolerance to noise in resource level measurements.

Adaptive Resource Modelling

19

IW PSS, October 2006


Thank you for listening.
Andrew Carrel University of Surrey a.carrel@surrey.ac.uk