From Streams of Observations to Knowledge-Level Productive Predictions

From Streams of Observations to Knowledge-Level Productive Predictions

Mark Wernsdorfer (University of Bamberg, Germany) and Ute Schmid (University of Bamberg, Germany)
DOI: 10.4018/978-1-4666-3682-8.ch013
OnDemand PDF Download:
$30.00
List Price: $37.50

Abstract

The benefit to be gained by Ambient Assisted Living (AAL) systems depends heavily on the successful recognition of human intentions. Important indicators for specific intentions are behavior and situational context. Once a sequence of actions can be associated with a specific intention, assistance may be provided by anticipating the next individual step and supporting the human in its execution. The authors present a combination of Sequence Abstraction Networks (SAN) and IGOR to guarantee early and impartial predictions with a powerful detection for symbolic regularities. They first generate a hierarchy of abstract action sequences, where individual contexts represent subgoals or minor intentions. Afterwards, they enrich this hierarchy by recursive induction. An example scenario is presented where a table needs to be set for several guests. It turns out that correct predictions can be made while still executing the observed sequence for the first time. Support can therefore be completely individual to the person being assisted but nonetheless be very dynamic and quick in anticipating the next steps.
Chapter Preview
Top

1. Introduction

Research in Ambient Assisted Living (AAL) aims to provide technologies for supporting people in their everyday activities (Wilken, Hein, & Spehr, 2011). For effective support, an automated assisting system needs to make valid assumptions fast and with a low number of initial examples. Furthermore, the system should be tolerant for individual idiosyncrasies, that is, it should not discourage people from their personal way of everyday interaction. Especially for elderly and impaired persons, it is important that the system maintains as much personal autonomy as possible to allow for as much cognitive and physical activity as possible (Schmid, 2008). For these reasons, individual behavior needs to be observed and no beforehand generalization over a large sample of persons may be imposed. Consequently, the challenge in creating a beneficial AAL system is combining the need for early predictions with the obligation to take each human into account individually.

One possible approach to design such a system is to infer the current intention (Stein & Schlieder, 2004; Kiefer & Schlieder, 2007) of a person by monitoring her/his activities and to predict a future intended sequence of actions on that basis. Typically it is not possible to identify the intention of an agent based on an isolated activity. For example, when a person raises from the living room chair, many consecutive actions are possible, dependent on the underlying intention. If, however, raising from the chair is motivated by the intention to get something to drink, the set of following activities are obviously reduced—walking to the kitchen is the most probable one.

In many Activities of Daily Living (ADL), we perform repetitive sequences of actions. For example, when drying dishes, we might dry three plates or ten plates, depending on the number of people served. Typically, the drying of each plate follows the same action pattern—take out of the sink with left hand, take the towel with the right hand, wipe the dish, put it away. The challenge for such types of sequences is to be able to predict the complete course of actions correctly for arbitrary situations. That is, the intention “dry dishes” is realized by action sequences of different length, depending on the number of dishes.

Rules which can capture the recursive character of such action sequences are called productive (Chomsky, 1965). To predict sequences with arbitrary length, it is necessary to learn the underlying set of productive rules, i.e., recursive function. The most prominent domain of productive rule learning is language acquisition (Chomsky, 1965; Marcus, 2001; Tomasello, 2003). Children learn the grammar rules underlying their mother tongue from positive experience. The learning of productive rule sets for generation and application of regular action sequences can be seen as analogous to the grammar acquisition problem. The challenge for an AAL system is to generalize productive rule sets from observed action sequences assessed by behavior monitoring and to predict succeeding action sequences of arbitrary length based on inferred intentions.

Research in the context of AAL currently focuses on identification of actions but not on prediction of action sequences (Wilken et al., 2011; Busch, Witthöft, Kujath, & Welge, 2011; Schröder, Wabnik, Hengel, & Goetze, 2011; Nesselrath, Lu, Schulz, Frey, & Alexandersson, 2011). We propose to improve the scope of AAL systems by implementing predictive capabilities. Definite assistance can be offered if intention recognition helps improving action anticipation.

Although San is a system that is being developed with autonomous agent control in mind, yet one of its major aims is to represent and recognize different sequences of discrete events. The goal is to replicate human unconscious anticipation in every-day interactions with the world. This serves two purposes in the case of an AAL system. The first one is differentiating types of sequences into high-level intentions. The second one is reliable anticipation of successive events with a very low number of initial examples. Using the sequence type information obtained by San, Igor is able to generate a recursive rule set that describes the observed sequences. Thus Igor is able to capture even complex regularities within sequences. Eventually, the recursive rule set allows for productive anticipation of successive events.

Complete Chapter List

Search this Book:
Reset