Efficient Incremental Plan Recognition method for Cognitive Assistance Hamdi Aloulou1, Mohamed Ali Feki2, Clifton Phua3 and Jit Biswas2 1
Ecole Nationale d’Ingenieure de Sfax Networking Protocols Department, Institute for Infocomm Research, Singapore, 3 Data Mining Department, Institute for Infocomm Research, Singapore
2
{mafeki}@i2r.a-star.edu.sg Abstract. In this paper we propose an efficient and incremental plan recognition method for cognitive assistance. We design our unique method based on graph matching and heuristic chaining rules in order to deal with interleaved and sequential activities. The finding of this research work is to be applied to predict abnormal behavior of the users, and optimize assistance for them. We have studied a use case of kitchen environment during lunch time that we will discuss in this paper and targeted Dementia patients. We will present implementation details as well as our evaluation plan. Keywords: Dementia assistance, plan recognition
Introduction: With the increasing number of aging population and the need to reduce hospital care costs, there is an emerging need from governments worldwide to invest in home care in order to come up with new technological solutions allowing elderly people to stay as long as possible in their homes. People with Dementia, as one category of people suffering from cognitive impairments, represents 5% of all persons above 65 and over 40% of people over 90 years [1][2]. Their care costs are greater than other elder’s costs and they require specially trained carers [3]. The home care market attracts many companies and corporations to invest in research and development. One of the main and important reasons that patients with Dementia are moved to the hospital is their inability to perform their Activities of Daily Life (ADLs) and Instrumental Activities of Daily Life (IADLs) independently. To address this issue, we are engaged in work with local nursing homes and hospitals in order to deploy assistive solutions able to help people with Dementia navigate their day independently. This work is carried out as a part of a research project in collaboration with Alexandra hospital, Singapore to assist people suffering from Dementia. The necessity for continuous 24/7 supervision of patients suffering from this type of disease has given birth to this collaboration with the goal of providing a platform that is capable of monitoring important activities of these patients and assist them to maintain their autonomy.
State of the Art In our application, we target online activity and plan recognition. Several works address the problem of plan recognition using probabilistic approaches [4][5][6], logic
based methods [7][8][9] or learning techniques [10][11][12][13][14]. However, they omit online recognition and base their experimentation on simulated environment. In our work we adopt an application driven approach with online plan recognition. We start our experimentation by exploring the weighted graph theory as described in the next section.
Implementation Details of our proposed solution To address the real needs raised by the specialist and described in the last section, we have prepared a laboratory environment with ambient sensors, that represent a ‘smart’ kitchen environment of an elderly person living alone. Our system aims to cover and detect cognitive errors as pointed out earlier and to redirect patient to the right activity plan. As a starting point, we assume that doctors and carers plan all activities that must be realized by the patient. We use graphs as an abstract representation of all activities fixed by doctors and carers. The vertices of a graph represent activities that patient performs in a specific scene, and edges are weighted by the maximum time between two successive activities. The application receives information sent from different wireless and wired sensors spread out in the scene. The sensors are mostly ambient, however we also consider simple non-intrusive wearable sensors. These sensors gather information that help us to detect all activities performed by the patient. When a patient enters a scene, a graph of activities is created (called the patient activity graph (PAG)). Each activity that the patient performs is detected by sensors and sent to the application which verifies the existence of this activity in the scene graph: If the activity doesn’t exist in the scene graph, an alarm is triggered to indicate an abnormal activity of the patient. If the activity exists in the scene graph, then the application verifies the existence of this activity in the PAG, and checks the following steps: If it is contained in this graph and the time has exceeded the maximum time for the activity, then the application concludes that the patient has entered into a completion error, and proposes to him to perform the activity that follows the last activity accepted by the application if it isn’t contained in the PAG, otherwise the application informs the patient that the activity plan is achieved. If the activity isn’t contained in the PAG, then the application verifies the existence in the PAG of one of activities that precedes this activity in the scene graph. If it is the case, then the application adds the activity performed, to the PAG and triggers a timer with the duration separating current activity and activities that follow it to monitor and control completion error and realization error. When the timer duration has expired, the application requests the patient to perform the activity that follows the activity for which the timer was launched, and restarts the timer. If the activity is performed before the end of the timer duration, then it is stopped. If no preceding activity is found in the scene graph, the application concludes that the patient has realized a judgment error, sequence error or organization error , he has jumped some steps in his activities plan. The application searches for the last activity performed in this activities plan and requests the patient to perform the appropriate activity that follows it.
If the patient leaves a branch of the graph to a new branch, then a timer is launched with an appropriate duration to control realization errors. When the timer duration has expired, the application requests the patient to perform the appropriate activity that follows the activity that was previously left, and restarts the timer. If he performs this activity before the end of the timer duration, then it is stopped. To control Initiation errors, the application monitors and controls the timing of lunch and medication. When it is time, it requests the patient to perform appropriate activities plan, and launches a timer with a considered duration. The timer is stopped when the patient performs the activity, otherwise, after the timer duration has expired, the application requests the patient to perform the activities plan once more and launches the timer once again. The description of the implementation details are depicted in the following class diagram.
Figure 1: Activity and Scene graphs main classes
Conclusion: In this paper we presented a novel approach based weighted graph matching of recognized activities to the Scene Graph and the Patient Activity Graph. We motivated our method by a real study and use case scenario provided by medical specialist. We will be able to provide first results of our prototype during the conference main event. References: [1]Fratiglioni L, Launer LJ, Anderson K, et al. Incidence of dementia and major subtypes in Europe: A collaborative study of population-based cohorts. Neurol Dis Elderly Res Group. Neurology,2000; 54:10-5.
[2]Launer LJ, Hofman A. Frequency and impact of neurologic diseases in the elderly of Europe: a collaborative study of population based cohorts. Neurology, 2000; 54:1-3. [3] (http://www.researchandmarkets.com/reports/337373/337373.htm) [4] Albrecht D.W., Zukerman I., Nicholson A.: Bayesian Models for Keyhole Plan Recognition in an Adventure Game, User Modelling and User-Adapted Interaction, (1998), (8) 5-47. [5] Boger J., Poupart P., Hoey J., Boutilier C., Fernie G. and Mihailidis A.: A Decision-Theoretic Approach to Task Assistance for Persons with Dementia. In Proc. of the International Joint Conference on Artificial Intelligence (IJCAI’05), Edinburgh, Scotland, (2005), 1293-1299. [6] Charniak E., Goldman R.: A Bayesian Model of Plan Recognition, Artificial Intelligence Journal, (1993), (64) 53-79. [7] Camilleri G.: A Generic Formal Plan Recognition Theory, In: IEEE International Conference on Information, Intelligence and Systems ICIIS’99, (1999), 540-547. [8] Nerzic P.: Two Methods for Recognizing Erroneous Plans in Human-Machine Dialogue, In: AAAI 1996 Workshop: Detecting, Repairing and Preventing HumanMachine Miscommunication, (1996). [9]Wobke W.: Two Logical Theories of Plan Recognition, Journal of Logic Computation, Vol. 12 (3), (2002), 371-412. [10] Bain M. and Sammut C: A framework for behavioral cloning, in S. Muggleton, K. Furakawa and D. Michie (Eds.), Machine Intelligence 15, Oxford University Press, 1995. [11] Bauchet J., Mayers A.: Modelisation of ADL in its Environment for Cognitive Assistance, In: Proc. of the 3rd International Conference on Smart homes and health Telematics, ICOST’05, Sherbrooke, Canada, (2005), 221-228. [12] Lent M., Laird J.E.: Learning procedural knowledge through observation, Technical Report, University of Southern California, ACM press, New York, NY, USA, (2001). [13] Liao L., Fox D., Kautz H: Learning and Inferring Transportation Routines. In Proc. of the National Conference on Artificial Intelligence (AAAI-04), San Jose, CA, (2004), 348353. [14] Wilson D.H. and Philipose M.: Maximum A Posteriori Path Estimation with Input Trace Perturbation: Algorithms and Application to Credible Rating of Human Routines, In Proc.