Matching SmartHouse Technology to Needs of the Disabled and Elderly Genevieve Davies, Nirmalie Wiratunga, Bruce Taylor and Susan Craw THE ROBERT GORDON UNIVERSITY ABERDEEN School of Computing Scott Sutherland School
Overview SmartHouse Technology Case Representation Iterative Retrieval Evaluation
SmartHouse Technology Humidity sensor Light IR Pressure Mat Remote Control Wall Switch Thermostat Switch Water Level Windows Open Window Sensor Access Ramp Basin Bath Blinds Clock Display Timer Community Alarm Cooker Curtains Dishwasher Extractor Fan Fridge Freezer Light Medication Dispenser Microwave Oven PC Entry Phone Radiator Sink Telephone Stereo TV Washing Machine WC
SmartHouse Textual Cases... Ms M is an indoor wheelchair user with cerebral palsy. She is a tenant in her own ground-floor flat, and whilst living moderately independently, did have support from care workers to assist her in getting dressed and bathed... She required an intercom that was both hands free and with video so that she could operate it from her wheel chair... also electrically operated locks were fitted on her external door... wheel-chair- indoor wheel-chair- outdoor house-typecare-staffable-into- bath Yesunknownground- floor-flat Yeswith-help Electrically-operated- locks-external-door Intercom-front-door YesVideo-hands-free
Case Representation Problem space: person description FieldBusArchitecture (1) Solution space: SmartHouse Devices
Retrieval With a DT Index Case Base C4.5 Decision Tree Index k-Nearest Neighbour Similarity Matching progress of retrieval Vote Tcl adaptation rules witch $idx { 0 {ProcBathroomAdapt()} 1 {ProcControllerAdapt()} 2 {ProcCookerDetectorAdapt()} 3 {ProcDoorAdapt()} Tcl iterative retrieval for {set i 0} {$i < $Counter} {incr i} { ProcMultipleConcepts() $i ProcRetrieveCases() $i ProcRetrieveAttributes() $i ProcWeightedAttributes() $i ProcAskUser() $i }... Set concept
Retrieval Example Insufficient cases for one shot retrieval Deal with each device separately Y (1 case, 50%), N (1 case, 50%)
Task Decomposition FieldBusArchitecture (1) Solution space: SmartHouse Devices
Generalised Concept CookerRelated Devices CaseX: ValueCaseY: Value auto-cooker-shut-off{gas, heat, timer}N detectors-over-cooker{gas, heat, smoke}unknown alarm-for-carersNN cooker-isolation-alarmNN fire-alarmYN gas-alarmYN FNeededNot-Needed Summarise each device group / sub-task Create index tree for each sub-task
Iterative Retrieval P = probe case CB = case base device-groups = 14 FOREACH sub-task {1 TO device-groups} concept <= generalised-concept (sub-task) tree <= build-tree (concept, CB) neighbours <= traverse (tree, P) k-sim-cases <= kNN (neighbours, P, sim-metric) sub-solution <= majority-solution (k-sim-cases) END FOREACH
Evaluation Leave-one-out testing –23 cases ( provided by domain expert) –real case used separately as probe to system –22 cases in case base User Testing –cases created by expert –A and B fewer device requirements –C more challenging Retrieval Strategies –Single, one-shot –Multiple, iterative
Quality of System Solution Doors Device Group SolutionCategories SystemExpertTPTNFPFN alarm-when-openNo 0100 contacts-on-external-doorsYesNo0010 contacts-on-bedroom-doorNo 0100 contacts-on-bathroom-doorNo 0100 contacts-on-kitchen-doorNo 0100 intercom-to-front-doorAudioVideo-HF0010 powered-external-doorsYes 1000 electric-locks-external-doorsNoYes0001 Overlap Sim = (TP+TN)/(TP+TN+FP+FN) Recall (R)=TP/ (TP+FN) Precision (P)= TP / (TP+FP) Fmeasure (F)= 2PR/(P+R)
Leave-One-Out Testing Multiple –improvement in 8 of the 14 device groups –overall solution quality Sim 0.91% F 0.64% Interesting alternatives suggested –flood detectors vs. push-to-operate –pendant activated community alarm vs. pager Overly expensive suggestions –environment and lighting controllers –person with some mobility problems
Device Group No Of Devices SingleMultiple SimF F Bathroom Controllers Cooker Detectors Doors Hearing Aids Movement Detectors Phone Related Remote Alarms Security Taps Technology Architecture Visual Aids Water Detectors Windows Overall Average
User Testing Patient A and B –2 out of 3 devices suggested for A –failure with controllers due to sparse case base –all 5 devices identified for B –decreased solution quality with increased neighbourhood –but Single achieves similar results Patient C more challenging –7 of the 9 devices identified by Multiple, compared to 2 –Multiple reused solutions from 9 different cases –R drops from 0.64 to 0.18 with Single
Conclusions Task decomposition is easy when there are fewer interactions between sub-tasks. Standard DT-based iterative retrieval strategy –focus on a sub-task at a time –addition of generalised concept Best use of case base with multiple indices Iterative retrieval provided better quality solutions compared to one-shot retrieval –empirical evaluation –subjective evaluation Use of multiple cases to solve different parts of a problem case was well received
Device Test Case ATest Case B group TPTP TNTN FPFP FNFN TPTP TNTN FPFP FNFN Bathroom Controllers Cooker Detectors Doors Hearing Aids Movement Detectors Phone Related Remote Alarms Security Taps Technology Architecture Visual Aids Water Detectors Windows Total
Device group SINGLEMULTIPLE TPTNFPFNTPTNFPFN Bathroom Controllers Cooker Detectors Doors Hearing Aids Movement Detectors Phone Related Remote Alarms Security Taps Technology Architecture Visual Aids Water Detectors Windows Total Recall Precision Sim 0.77