Synthetic Elderly Companions

For decades, we have been promised the personal service of artificial agents—entities able to perform services and tasks on our behalf. Science fiction is replete with examples of loyal and dutiful artificial helpmates. Here, we choose the elderly companion domain because it is a challenging application in much need of attention. According to the World Health Organization, the world's population over 60 years of age will double to exceed 2 billion by the year 2050 (WHO, 2017). A significant percentage of elderly people suffer from dementia, depression, anxiety, and substance abuse. As we get older, we naturally encounter cognitive and physical degradation often requiring the assistance of another person. Furthermore, a study of 6,500 elderly men and women showed a lack of social contact leads to an early death, regardless of the presence or absence of underlying health issues (Steptoe et al., 2013). Elderly health issues such as heart disease, diabetes, high blood pressure, and high cholesterol contribute to complications with declining vision, hearing and motor skills.

It is common practice for an elderly person to move in with younger family members or to hire a caregiver to visit with the elderly person on a daily basis. However, as the number of elders grows, there will likely not be enough skilled caregivers. There is a tremendous need for artificial and synthetic caregivers for the elderly. In the United States alone, elder care was projected to be worth approximately $400 billion by 2018 with in-home healthcare services is the second largest and fastest growing segment (Buitron, 2017).

The requirements of elder care are extensive. Building a synthetic expertise for elder care pushes and tests the state of the art in cognitive systems and artificial intelligence. At the heart of a synthetic elderly companion is a piece of software able to assist the elder—a softzvare agent. The field of software agents can be traced to Hewitt's Actor Model describing a self-contained, interactive and concurrently-executing object, possessing internal state and communication capability (Hewitt et al., 1973; Hewitt, 1977). The idea of autonomous, goal-driven software agents has evolved from multi-agent systems (MAS), distributed artificial intelligence (DAI), distributed problem solving (DPS), and parallel artificial intelligence (PAI). Because of their independent nature, software agents promise modularity, speed (due to parallelism), reliability (due to redundancy), knowledge level description, easier maintenance, reusability, and platform independence (Huhns and Singh, 1994). Fundamentally, a software agent is:

software and/or hardware capable of acting del iteratively to accomplish

tasks on behalf of its user.

King (1995) identifies several different kinds of software agents: search agents, report agents, presentation agents, navigation agents, role-playing agents, management agents, search and retrieval agents, domain-specific agents, development agents, analysis and design agents, testing agents, packaging agents, and help agents. Nwana (1996) identifies the following types of software agents:

• Collaborative Agents agents cooperating with other agents to

perform tasks

• Interface Agents personal assistants in collaboration with a

user

• Mobile Agents capable of roaming networks (such as the

Internet)

• Information Agents managing/collating information from

distributed sources

• Reactive Agents lacking internal knowledge representation/

reasoning

• • Hybrid Agents a combination of two or more agent types
• • Heterogeneous Agents a collection of different agent types
• • Smart Agents agents capable of human-level cognition

When these types of agents were identified in the mid-1990s the idea an agent could perform high-level cognition (smart agents) seemed far in the future. Now, some 25 years later, this future has arrived. The synthetic elderly companion we envision in this chapter is a smart interface agent capable of collaboration, mobility, and information acquisition and human-level processing. We view the synthetic elderly companion as not only a personal assistant, but also a friend and confidant for the elder.

Some virtual home assistants are on the market now or are in development, however, they are limited in ability and robustness. Catalia Health's Mabu is designed to be a personal healthcare companion with the ability to socially interact and assist patients with the medication portion of their treatment (Kidd, 2015; Catalia Health, 2019). Intuition Robotics' ElliQ is aimed at keeping older adults active and engaged by connecting them to their families and the outside world (Elliq, 2019). ElliQ is a friendly, intelligent, inquisitive presence in the elder's daily life able to offer tips and advice, respond to questions, and surprise with suggestions. Asia Robotics' Dinsow is a service robot designed for elderly care service (Dinsow, 2019). Riken's Robobear is an experimental nursing care robot capable of performing tasks such as lifting a patient from a bed into a wheelchair or providing assistance to a patient who is able to stand up but requires help to do so (Riken, 2015). Among other systems in various stages of development are: Pillo, Aido, Jibo, and Oily (Inventions World, 2018). Like Mabu, these are physically small, partially mobile, figurines with an expressive human-like face and some sort of small display screen (e.g., a tablet) for interaction. All of these systems feature natural language speech recognition and synthesis. While each of these systems performs a task, they are not comprehensive elderly caregivers.

Our goal in this chapter is to define a synthetic elderly companion based on our Model of Expertise shown in Chapter 7. We call our elderly companion Lois (Loved One's Information System). Physically, Lois is embodied in a number of display screens, microphones, speakers, and cameras located throughout the home of the elder. Logically, Lois is a cog with which the elder works with and relies on as an assistant. Together, the elder and Lois form a human/cog ensemble—a synthetic elderly companion. The elderly population should be willing to adopt Lois and related technology. The Pew Research Center reports 67% of adults over the age of 65 are Internet users and go online (Anderson and Perrin, 2017). Seniors are also using smartphones, computers, and tablets at increasing rates.

As with a human caretaker, critical for Lois is to maintain situational awareness and contextual awareness of the elder throughout the day. Lois must be able to determine when the elder is sleeping, napping, eating, exercising, etc. and monitor the elder's overall well-being. Figure 10-1 shows the knowledge-level and expertise-level architecture of Lois. Overall, Lois must continually determine the status of the elder in several aspects:

• 1. Health overall health including vital statistics
• 2. Mood emotional state
• 3. Cognitive mental state/acuity, clarity, coherence, attention
• 4. Sleep quality, duration
• 5. Meals food/beverage intake, calories, fat, sugar, carbohydrates
• 6. Meds type, time, quantity
• 7. Hygiene grooming, bathing
• 8. Activity general activity level, kinetics, movement, excursions
• 9. Social interaction with friends/others, online, visitors
• 10. Exercise structured physical therapy, type, quantity/duration

Fig. 10-1: Elderly companion (Lois).

Elder Models

To monitor the elder's status, Lois maintains a primary model of the elder (M_elder). This model contains up-to-date general information about the elder (e.g., name, age, gender, family, address, telephone number, email addresses, social media logins, automobile, license, insurance information, emergency contacts, current location, etc.)- Since models are both hierarchical and associative in nature, the primary model is linked to a sub-model for each of the aspects listed above:

Each of these models contain information relevant to the aspect, exemplar information, target goals, and current values for the elder. Through a variety of sensors, Lois continually monitors the elder and the elder's environment, via the perceive function, to generate T, the perceived state. Based on its perceptions, Lois updates the elder model (M_elder) and the relevant sub-models. This way, Lois can determine the elder's current state and monitor the activities the elder is currently engaged in.

A collection of models for each family member is maintained as well (M_famil) allowing Lois to recognize each family member and tailor interaction with that family member based on legal, ethical, and privacy settings (e.g., HPPA-approved family list).

Activity Models

To monitor the elder's activities, Lois maintains a family of activity and behavior models (M_activity). Examples of activity models are:

Activity models contain descriptions of each activity allowing Lois to recognize the activity based on its observations. Since models are dynamic data stores, over time, Lois is able to capture idiosyncrasies of the specific elder being cared for by continually updating the activity models using the evaluation and analyze skills. Also, Lois will be able to learn the elder's routines allowing Lois to calculate expectations such as "on a weekday, the elder wakes about 8 AM."

Each activity model represents a unique behavioral context allowing Lois to tailor interaction with the elder accordingly. For example, Lois may structure dialog with the elder differently when the elder is dressing as opposed to when the elder is reading. Learning routine and idiosyncrasies allows Lois to detect departures from normal behavior, a key skill for elderly companions.

Emotional/Mood Models

Likewise, Lois maintains a family of emotional and mood models (M_emoljon) allowing Lois to monitor the elder's emotional state. Sample emotional models are:

It is important to note behavioral and emotional models represent states the elder can be in at any point of the day and the elder may be in multiple states simultaneously. For example, the elder could be eating, reading, and watching television all at the same time. Likewise, the elder could be lonely, worried, and restless at the same time. Activity, emotional, and elder models are all domain-specific models so are subsets of M_D. The evaluate skill allows Lois to judge the elder's emotional and mental state.