deploying raxem2 planning improvements in daily work practice

31
Deploying RAXEM2 Planning Improvements in Daily Work Practice Giulio Bernardi , Amedeo Cesta & Gabriella Cortellessa ISTC-CNR [PST] Institute for Cognitive Science and Technology National Research Council of Italy Planning and Scheduling Team http://pst.istc.cnr.it Work supported by ESA - the European Space Agency

Upload: lel

Post on 16-Jan-2016

29 views

Category:

Documents


1 download

DESCRIPTION

Deploying RAXEM2 Planning Improvements in Daily Work Practice. Giulio Bernardi , Amedeo Cesta & Gabriella Cortellessa ISTC-CNR [PST] Institute for Cognitive Science and Technology National Research Council of Italy Planning and Scheduling Team http://pst.istc.cnr.it. - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: Deploying RAXEM2 Planning Improvements in Daily Work Practice

Deploying RAXEM2Planning Improvements in Daily Work Practice

Giulio Bernardi, Amedeo Cesta & Gabriella Cortellessa

ISTC-CNR [PST]

Institute for Cognitive Science and Technology

National Research Council of Italy

Planning and Scheduling Team

http://pst.istc.cnr.it   

Work supported by ESA - the European Space AgencyWork supported by ESA - the European Space Agency

Page 2: Deploying RAXEM2 Planning Improvements in Daily Work Practice

Introduction

• Raxem is a software system developed to plan the upload of commands to the Mars Express spacecraft (orbiting around Mars since 2004)

• Raxem project started after the success of Mexar2, which is used to plan data dumping activities from the spacecraft since 2005

Page 3: Deploying RAXEM2 Planning Improvements in Daily Work Practice

Introduction

MEXAR

RAXEM

Page 4: Deploying RAXEM2 Planning Improvements in Daily Work Practice

Commands from Earth

• The Mars Express probe receives instructions from Earth

• Instructions are time-tagged telecommands (TCs)

• On board, TCs are stored in a memory buffer, the Master Timeline (MTL)

• TCs are discarded after execution

• Deciding how to send data is not trivial…

Page 5: Deploying RAXEM2 Planning Improvements in Daily Work Practice

How to send commands?

• Related TCs are grouped in MTL Detailed Agenda Files (MDAFs)

• Analogous to programs: they must be uplinked in an unique pass

• Many constraints:– Communication is only possible during certain

uplink windows– Sending data to Mars takes time– The Master Timeline has a finite capacity– A contingency plan is needed in case of failure

Page 6: Deploying RAXEM2 Planning Improvements in Daily Work Practice

The Mex-Up problem

• An uplink plan must satisfy these goals– specifying when to uplink each MDAF– which modality of transmission should be

employed, and – identifying a secondary uplink window to use in case

of failures

All requested MDAFs must be uplinked as early as possible,so that they are on board in time, given the finite capacity of

the MTL and the limited bandwidth of the transmissionchannel, while trying to keep the MTL as full as possible

Page 7: Deploying RAXEM2 Planning Improvements in Daily Work Practice

The Mex-Up problem

MDAF files

Earth

Master Time Line (MTL)

AASFO1A3AASFO1AXAASFO1A3AASFO1A4AASFO1A3AASFO1A5AASFO1B3AASFO1B3AASFO1J3…

Telecommands (TCs)

Mars

RAXEM2

Spacon Instruction Form (SIF)

Uplink Windows

OWLT -- One-Way Light Time

Page 8: Deploying RAXEM2 Planning Improvements in Daily Work Practice

Before Raxem

• The complexity of the Mex-Up problem was initially underestimated

• Planning was carried on manually with paper and pencil

• Disadvantages of manual planning:– Very time-consuming– Prone to errors– Poor optimization– Emergency re-planning was extremely difficult

Page 9: Deploying RAXEM2 Planning Improvements in Daily Work Practice

The Raxem Solution

• A first version of Raxem was developed to assist users in solving the Mex-Up problem

• Notable features:– Interactive plan generation– Graphical output of the solution (charts)– Inspection of spacecraft status

• Raxem has been operational since summer 2007 [see Cesta et al ECAI-08]

• The success of Raxem encouraged users to ask for a more complete tool…

Page 10: Deploying RAXEM2 Planning Improvements in Daily Work Practice

Need for better integration

• A semi-manual procedure was employed to parse Raxem’s output and produce SIF forms (the official instructions delivered to uplink operators)

• History of performed uplinks was held by external means

• There was the opportunity to transform Raxem in a tool able to embrace the whole uplink work cycle

Page 11: Deploying RAXEM2 Planning Improvements in Daily Work Practice

More than a planning tool

• New goals for Raxem2:– Ability of maintaining the whole history of uplink

operations (uplink database)

– Acknowledgement of arbitrary changes on the planned solution by the users

• This is necessary to accurately record actual uplink operations

– SIF generation and management

– User management and graphical improvements

Page 12: Deploying RAXEM2 Planning Improvements in Daily Work Practice

• Raxem2 includes a new module that ensures complete and continuous management of the uplink problem

• Existing functionalities have been comprehensively enhanced

Uplink DataBaseUplink

DataBase

Domain Modeling

Problem Solving

AI Module

User Interaction

Spacon Instruction Form (SIF) Generator

Plan life cycle services

RAXEM2

Interaction Module

UserManagement

The Raxem2 solution

Page 13: Deploying RAXEM2 Planning Improvements in Daily Work Practice

Modeling with Timelines

• Two relevant components of the model are the MTL and the Communication Channel

• They are timelines:– They can report their status over time– Data can be allocated/deallocated on them over

time

• The MTL is a cumulative resource– Has a finite capacity

• The Channel is a binary resource– Is either busy or free in a particular instant

Page 14: Deploying RAXEM2 Planning Improvements in Daily Work Practice

Modeling with Timelines

• The act of data uplink is an Activity

• An activity is modeled by two operations:– Channel operation

• Represents the transmission of data during an uplink window

• Requires the whole bandwidth of the channel for the entire duration

– MTL operation• The act of storing data in the MTL• It “instantaneously” increases the amount of data in

the MTL

Page 15: Deploying RAXEM2 Planning Improvements in Daily Work Practice

Modeling with Timelines

1) MTLCumulative

resource

2) ChannelBinary

resource

Last(MDAF)

stUpLink

First(MDAF)

durUpLinkChannel operation

MTL operation

Size(MDAF)

Page 16: Deploying RAXEM2 Planning Improvements in Daily Work Practice

The Solver

• The software component that generates a plan

• It employs a relatively simple constructive algorithm which provides a good trade-off between speed and optimality of the solution

• It can relax constraints for some MDAFs if needed:– Confirmation Scheme (full or reduced?)– Uplink Scheme (secondary window or not?)

• The user can change them too by providing hints to the solver

Page 17: Deploying RAXEM2 Planning Improvements in Daily Work Practice

The main algorithm

foreach mdaf to uplink do while not allocateMDAF(mdaf,currentTime) do if can relax then relax(mdaf) else break end while if allocated(mdaf) then currentTime primaryUplinkEnd(mdaf)end for

function allocateMDAF(mdaf, currentTime) start firstAvailableInstant(currentTime) while not allocated(mdaf) do try if not multiMDAFAllocation(mdaf,start) then singleMDAFAllocation(mdaf,start) on RetryLaterError do start instantToRetry end try end whileend function

foreach mdaf to uplink do while not allocateMDAF(mdaf,currentTime) do if can relax then relax(mdaf) else break end while if allocated(mdaf) then currentTime primaryUplinkEnd(mdaf)end for

function allocateMDAF(mdaf, currentTime) start firstAvailableInstant(currentTime) while not allocated(mdaf) do try if not multiMDAFAllocation(mdaf,start) then singleMDAFAllocation(mdaf,start) on RetryLaterError do start instantToRetry end try end whileend function

Page 18: Deploying RAXEM2 Planning Improvements in Daily Work Practice

The Persistence Module

• Raxem’s data is held in the uplink database– It contains both the historical data and the working set

• Persistence is handled using DAO objects implementing CRUD methods

• The database backend is provided by SQLite library– All data is contained in a single file– No need for an external system– Backups and software deployment are easy

Uplink DataBaseUplink

DataBaseSQLiteSQLite uplink.dbuplink.db

Page 19: Deploying RAXEM2 Planning Improvements in Daily Work Practice

Interactive planning

• Raxem2 encourages user’s involvement in planning

• Planning parameters can be altered for individual MDAFs:– Use reduced confirmation (shorter uplink)– Don’t allocate a secondary window– Exclude some MDAFs from the plan

• Possibility of tuning the solution

• Allows to perform what-if analysis

Page 20: Deploying RAXEM2 Planning Improvements in Daily Work Practice

The Raxem2 interaction

MDAF view

Input for the AI solver– Confirmation scheme– Uplink scheme

SIF view

“Solve” button

Page 21: Deploying RAXEM2 Planning Improvements in Daily Work Practice

Great Flexibility

• Users can override any decision made by the solver

• Raxem2 accepts any change, and issues warnings if needed

• Users can also inform the system about uplinks that actually happened

• In both cases, the internal model of the problem is updated accordingly

Page 22: Deploying RAXEM2 Planning Improvements in Daily Work Practice

The SIF form view

Uplink intervals(can be overridden)

Comment areas

Page 23: Deploying RAXEM2 Planning Improvements in Daily Work Practice

Solution view: before Raxem

Olligram: hand-made gantt-like chart

Page 24: Deploying RAXEM2 Planning Improvements in Daily Work Practice

Initial solutions in Raxem v1

MTL Usage level

Uplink Graph

OnBoard Execution Graph

Page 25: Deploying RAXEM2 Planning Improvements in Daily Work Practice

The integrated view in Raxem2

MTL Usage level

Uplink Graph

OnBoard Execution Graph

Page 26: Deploying RAXEM2 Planning Improvements in Daily Work Practice

Experimental Results

• Accuracy and performance tests– Accuracy: degree of adherence of the solution

to user’s expectations (percentage)– Performance: time needed to produce a plan

(seconds)

• Two sets of data, about 60 MDAFs each

• Tests show program behavior with different planning parameters

• Raxem2 always outperforms Raxem by a great extent

Page 27: Deploying RAXEM2 Planning Improvements in Daily Work Practice

Experimental Results (2)

Accuracy

full w full w/0 reduced w/ reduced w/o

Set 1 93.18% 95.76% 92.73% 100.00%

Set 2 96.49% 100.00% 92.98% 100.00%

Performance (seconds)

full w full w/0 reduced w/ reduced w/o

Set1

Raxem2 0.755 0.387 0.731 0.316

Raxem 1.008 0.878 n/a 0.948

Improvement 33.51% 126.87% n/a 200.00%

Set2

Raxem2 0.814 0.360 0.784 0.352

Raxem 1.236 1.200 n/a 1.138

Improvement 51.84% 233.33% n/a 223.30%

Page 28: Deploying RAXEM2 Planning Improvements in Daily Work Practice

User Assessment

• Raxem2 is operational at ESA since March 2009

• Engineer work load was reduced from 5 hours to below 1 hour

• The software was up to expectations with regards to– plan generation– handling of emergency situations– flexibility

• SIF management has proven to be a very valuable feature

Page 29: Deploying RAXEM2 Planning Improvements in Daily Work Practice

Conclusions

• Raxem2 is a successful example of integration of different technologies

• Planning & Scheduling capabilities alone are not enough: users need a complete system to effectively perform their work

• Seamless integration with the established work practice is necessary for acceptance in a rather conservative environment

Page 30: Deploying RAXEM2 Planning Improvements in Daily Work Practice

Conclusions

• Key factors for success– Continuous user involvement in planning

• Small adjustments, what-if analysis

– Centrality of user in the decisional process• Not a “black box” system• Operators can override any choice

– Integration of different technologies• Needed to provide the most effective and complete

user services

Page 31: Deploying RAXEM2 Planning Improvements in Daily Work Practice

Questions

?