Artificial Intelligence in the Design of Assistive Technology Martha E. Pollack Computer Science & Engineering University of Michigan An Example of “Use-Inspired Research” The research is motivated by an important societal problem But is not overly constrained by the problem The research should have fundamental interest And thus be potentially useful for multiple applications The Population is Aging > Age 60 World 2000 2050 China 6.9% 22.7% 10.0% 21.4% India 7.5% 20.1% Japan 23.3% 42.4% Myanmar 6.8% 20.5% Belarus 19.3% 37.6% Germany 23.2% 34.5% Italy 24.1% 40.6% Australia 16.4% 29.9% Netherlands 18.2% 30.7% Fiji 5.7% 22.7% Slovenia 19.2% 41.5% Egypt 6.8% 18.7% United States 16.1% 25.5% Iran 6.4% 24.8% Mexico 6.9% 26.2% Jordan 4.6% 19.0% Brazil 7.8% 25.9% Columbia 6.9% 22.7% Botswana 4.2% 6.0% Ethiopia 4.6% 7.7% Mali 3.9% 5.3% Source: United Nations Population Division http://esa.un.org/unpp/ The “Oldest Old” > Age 80 2000 2050 World 1.1% 4.2% United States 3.2% 7.2% China 0.9% 7.2% Source: United Nations Population Division http://esa.un.org/unpp/ Intelligent Technology Can support people with Mobility impairment Obstacle-avoiding wheelchairs Declines in sensory function Courtesy, R. Simpson, Univ. Pittsburgh & AT Sciences Courtesy Oticon Adaptive digital hearing aids Social and emotional challenges Elder-friendly Internet and email systems Cognitive decline Courtesy Generations on Line The Challenge Cognitive impairment can impact performance of daily activities necessary for health and wellbeing ADLs: eating, dressing, bathing, toileting, . . . IADLs: managing medicines, housekeeping, arranging transportation, preparing meals, . . Can become difficult to follow a daily plan Autominder: An Adaptive Reminder System Uses Artificial Intelligence techniques to Model, update, and maintain the client’s plan Monitor the client’s performance Including complex temporal and causal constraints Updating the plan as execution proceeds Reason about what reminders to issue, and when Ensuring compliance, without sacrificing client independence Autominder Interaction Req/Opt R R O R Activity check blood glucose lunch TV check blood glucose Allowed 11:5012:10 12:151:00 17:0017:30 16:5517:15 Expected Observed 11:55 REMIND 12:42 12:40 REMIND 16:55 Autominder WhatArchitecture should the client do? Activity Info Plan Updates Plan Manager Client Modeler Sensor Data Inferred Activity Client Plan Client Model smart home Activity Info Preferences Client Model Info Intelligent Reminder Generator Reminders Plan Manager Maintains up-to-date record of client’s planned activities Techniques: AI Planning Temporal Constraint Satisfaction Plan Representation Plans are structured sets of activities causal connections temporal constraints qualitative, quantitative, disjunctive conditional constraints Temporal constraints modeled as Disjunctive Temporal Problems (DTPs) Disjunctive Temporal Problems A set of time points (variables) V and a set of constraints C of the form: lbji Xi – Xj ubji … lbmk Xk – Xm ubmk Solution: assignment of times to all variables, so that all constraints in C are satisfied Generalization of the Simple Temporal Problems Disjunctions critical to model non-overlap requirements Example Temporal Constraints Blood glucose should be checked between 11:50 and 12:10: C1: 11:50 BS – TR 12:10 Lunch takes between 15 and 30 minutes: C2: :15 LE – LS :30 The TV show takes 30 minutes: C3: :30 TVE – TVS :30 Watching TV can begin at 17:00 or at 18:00: C4: 17:00 TVS – TR 17:00 18:00 TVS – TR 18:00 Checking glucose and watching TV should not overlap: C5: 0 TVS – BE 0 BS – TVE Dishes should be taken within 1 hour of finishing lunch. C6: 0 DS – LE 60 DTP Solving as Meta-Level Propagation C1 : {c11 : y – x 5} C2 : {c21 : w – y 5} {c22 : x – y -10} {c23 : z – y 5} C3 : {c31 : y – w -10} NOT: c21, c31 NOT: c11, c22 Component STP: C1 c11, C2 c23 , C3 c31 One exact solution: {x = 0, y = 1, z = 2, w = 12} Temporal Constraint Satisfaction Efficiently solving DTPs Add temporal uncertainty [Tsamardinos & Pollack, AIJ 2003] [Morris, Muscettola, & Vidal, IJCAI 2001, Venable & Yorke-Smith, IJCAI 2005, Rossi, Venable, & Smith, CP 2004] causal uncertainty [Tsamardinos & Pollack, Constraints, 2003] preference functions [Khatib et al., IJCAI 2001, 2003; Peintner & Pollack, AAAI 2004, 2205, Morris, Morris, Khatib & Yorke-Smith, ICAPS Wkshp. 2005, Sheini, Peintner, Sakallah & Pollack, CP 2005, Moffitt & Pollack, AAAI 2006] partial satisfaction dynamic structure hybrid constraints [Moffitt & Pollack, FLAIRS 2005, Peintner, Moffitt, & Pollack, ICAPS 2005, Moffitt & Pollack, IJCAI 2005, Liffiton, Moffitt, Pollack, & Sakallah, IJCAI 2005] [Schwartz & Pollack, ICAPS Wkshp. 2005] [Moffitt, Peintner, & Pollack, AAAI 2005] Efficiently Solving DTPs Example: Removal of Subsumed Variables If this assignment to Ci is implied by the partial assignment above it, prune the other values for Ci Ci cij Ci cik Ci cil DTPs with Preferences Plan includes exercising and a visit from a friend Should finish exercise before visit or start after visit: VS – EE [5,] v ES – VE [0,] 3 2 1 not allowed V 0 5 10 15 … VS – EE 3 2 1 not allowed 0 5 10 15 … ES – VE Now an optimization problem Solution Technique [AAAI 2006] Convert each constraint into a set of valued constraints, one per preference level 3 2 1 not allowed V 0 5 10 15 … VS – EE 3 2 1 not allowed V=1 0 5 10 15 … ES – VE V=2 V=3 V 0 5 10 15 … VS – EE 0 5 10 15 … VS – EE 0 5 10 15 … ES – VE 0 5 10 15 … VS – EE Apply branch-and-bound with partial satisfaction Performs very well in comparison with previous methods Same Techniques Apply Elsewhere Rectangle Packing Rectangle i must be contained within the enclosing space of dimensions W x H: xi 0, yi 0, xi + wi W, yi + hi H Rectangle i and j must not overlap: xi + wi xj xj + wj xi yi + hi yj yj + hj yi Orientation of rectangles adds hybrid constraints: xi + B(oi) xj xj + B(oj) xi … Comparison of Approaches [ICAPS 2006] For the case of optimal packings fixed orientations Number of squares 12 16 19 State-of-the-art in… VLSI / Floorplanning1 7:45:49 TIMES OUT TIMES OUT Operations Research2 0:00:13 0:10:05 3:00:53:18 Artificial Intelligence 0:00:00 0:00:03 0:00:04:09 10 day time-out limit 1 2 (Chan and Markov, 2004) (Clautiaux, Carlier, & Moukrim, 2004) Autominder Architecture What has the client done? Activity Info Plan Updates Plan Manager Client Modeler Sensor Data Inferred Activity Client Plan Client Model smart home Activity Info Preferences Client Model Info Intelligent Reminder Generator Reminders Client Modeler Given what can be observed (sensor input, clock time, stored plan, reminder information, etc.), infer probabilities that various actions were performed Techniques: Wireless sensor networks Reasoning under uncertainty Ensuring privacy and security of data collected is paramount! Probabilistic Reasoning for Activity Recognition Use Hidden Markov Models (HMMs), Dynamic Bayes Nets (DBNs), etc. Example: Train one HMM per activity type; observed variables = sensors firings For discrete activities, compute probability of each type Key question: segmentation Make Tea: Pr(faucet) = .6 Pr(faucet) = .5 Pr(cup) = .4 Pr(tea) = .45 ... ... Courtesy M. Philipose et al., Intel Research Same Techniques Apply Elsewhere Recent work at Intel Seattle on using these techniques to train anesthesiologists Autominder Architecture Activity Info Plan Updates Plan Manager Client Modeler Sensor Data Inferred Activity What action should the Client system Plan take? Client Model Activity Info Preferences smart home Client Model Info Intelligent Reminder Generator Reminders Intelligent Reminder Generation Given a client’s plan and its execution status: Easy to generate reminders at earliest possible time of each action Harder to “remind well” Maximize likelihood of appropriate performance of activities Allow flexibility Facilitate efficient performance Avoid annoying client and/or making the client overly reliant Techniques: iterative refinement (local search) machine learning One Approach: Iterative Refinement [AIPS 2002] 8:00 12:00 B L 16:00 D 12:00 Midnight TV 8:30 12:32 8:00 12:00 B L 16:00 D 12:00 Midnight 8:00 12:00 B L 12:00 Midnight TV TV 16:00 D Alternative Approach: Learning [ICML 2004] Use reinforcement learning to induce the best interaction strategy Decide whether, when, and how to remind, given information about the client’s “state” Add a dynamic action proposer—the plan manager—to a standard RL architecture Augmented RL Architecture Environment Client Sensors Sensors Actuators Reminder Production Payoff Agent Adaptive Reminder Generator Selected Actions Percepts State Estimator Client Modeler Estimated State Plan Action Proposer Plan Manager Sample Experiment (with a client simulator) Autominder Platforms Key Research Challenges Integrated activity recognition—inside and outside the home Enhanced privacy assurance Elaboration of machine learning techniques for interaction strategies Platform and interface design Use of adaptive systems for additional purposes Social integration Cueing for behavioral adaptation Incoporation of other assistance mechanisms (e.g., face recognition) Field testing and iterative design!