KLEM: A Method for Predicting User Interaction Time and System Energy Consumption during Application Design CHRISTIAN DZIUBA ILYAS DASKAYA Ubiquitious Computing WS-2009 TU-BRAUNSCHWEIG Lu Luo and Daniel P. Siewiorek /07©2007 IEEE.

Intro  Motivation and Goals  KLM in Brief  KLEM – an overview  Modelling Process  Energy Characterizing Process  Interaction Benchmarks  An Experiment Setup  Results & Model Verification  Conclusion 09/04/2015 DZIUBA & DASKAYA UBI-COMP WS-2009 TU-BRAUNSCHWEIG 2

Motivation & Goals  Energy consumption from user interaction – Wide variety of user interaction methods – Highly interactive tasks – Unavailability of application at design time – High cost to conduct user study  Keystroke-Level Energy Model (KLEM) – Keystroke-Level Model (KLM) extended – Predicts expert user task time – Predicts system energy consumption of task 09/04/2015 DZIUBA & DASKAYA UBI-COMP WS-2009 TU-BRAUNSCHWEIG 3

What is “KLM”? (Keystroke Level Model) 09/04/2015 DZIUBA & DASKAYA UBI-COMP WS-2009 TU-BRAUNSCHWEIG 4  Created by Card, Moron, and Newell in 1980  Based on GOMS (goals, operators, methods, selection rules)  Describe a task by placing operators in a sequence K –keystroke P –point with mouse H –homing (move hand from mouse to keyboard) D (takes parameters) –drawing R (takes parameters) –system response time M –mental preparation  There are heuristic rules to insert candidate Ms into the sequence  Task execution time = Σall operators involved  KLM for pen-based handheld user interfaces Physical Operators Mental Operator

KLEM as an extension to KLM 09/04/2015 DZIUBA & DASKAYA UBI-COMP WS-2009 TU-BRAUNSCHWEIG 5  Given: a task, the methods to execute the task, the design, and a target platform  Predict: user time and task energy

KLEM – Modelling Process 09/04/2015 DZIUBA & DASKAYA UBI-COMP WS-2009 TU-BRAUNSCHWEIG 6  Using CogTool  Support touch-screen and voice-based interfaces  Create KLM by demonstrating task on storyboards  Sample model trace created by ACT-R

KLEM – Modelling Process 09/04/2015 DZIUBA & DASKAYA UBI-COMP WS-2009 TU-BRAUNSCHWEIG 7  We can define various types of widgets in CogTool which is supported by the respective hardware (eg. PalmTungsten 5)

Interaction Benchmarks 09/04/2015 DZIUBA & DASKAYA UBI-COMP WS-2009 TU-BRAUNSCHWEIG 8 After the user selects an item, the application records the selection, updates a small area of the display to look responsive to user operation, and goes back to the waiting/idling state for the next user operation. We define this kind of system activity “Small”. As for “Medium” and “Large” system activities, for example, if a “next” button is pressed to open a new window, the consequent system activity is defined “Medium”, while an “open” button that reads a 1MB file is considered “Large” system activity.

KLEM - Energy Characterizing Process 09/04/2015 DZIUBA & DASKAYA UBI-COMP WS-2009 TU-BRAUNSCHWEIG 9  Measurement based approach  Power state correspondence of KLM operators  “Busy”(K, D, R): MOTOR, WAIT-FOR-SYSTEM…  “Idle”(M): VISION, PROCEDUAL…  Operator time decided by KLM and benchmarks Button PressDisplay Update (more states can be added)

An Experiment Setup 09/04/2015 DZIUBA & DASKAYA UBI-COMP WS-2009 TU-BRAUNSCHWEIG 10  Platforms: Windows Mobile (iPaq) Palm OS (Tungsten) Method 1: map navigation interface Method 2,3,4: scroll list interface

Model Verification - user time 09/04/2015 DZIUBA & DASKAYA UBI-COMP WS-2009 TU-BRAUNSCHWEIG 11  User study: 10 participants Two platforms 12 tasks in total  Measurement: Total task execution time Prediction error:5.6% for iPaq 8.8% for Tungsten

Model Verification – task energy 09/04/2015 DZIUBA & DASKAYA UBI-COMP WS-2009 TU-BRAUNSCHWEIG 12  Measurement: System energy consumption during task  Prediction error: 4.4% for iPaq 8.4% for Tungsten

Conclusion 09/04/2015 DZIUBA & DASKAYA UBI-COMP WS-2009 TU-BRAUNSCHWEIG 13  A methodology based on well-established HCI theory and practices is introduced to make design-time prediction on interactive task energy consumption  The energy efficiency of different user interaction methods on the same task is compared and analyzed. Author show when using different interaction modalities, the energy consumption can vary by a factor of three in achieving the same user goal. Future work is to extend KLEM to other interaction modalities such as speech recognition for input and speech synthesis and speech playback for output.