Paper2 - Technische Universität Darmstadt

Fitness Tracker or Digital Personal
Coach: How to Personalize Training
Benedikt Schmidt
Technische Universität
Darmstadt
Hochschulstrasse 10
64289 Darmstadt, Germany
benedikt.schmidt@
tk.informatik.tu-darmstadt.de
Rüdiger Eichin
SAP SE
Dietmar-Hopp-Allee 16
69190 Walldorf, Germany
[email protected]
Sebastian Benchea
SAP SE
Dietmar-Hopp-Allee 16
69190 Walldorf, Germany
[email protected]
Christian Meurisch
Technische Universität
Darmstadt
Hochschulstrasse 10
64289 Darmstadt, Germany
christian.meurisch@
tk.informatik.tu-darmstadt.de
Abstract
The use of activity tracking systems promises support in
meeting physical fitness goals. This support generally focuses on an improved self-awareness based on the review of own fitness data, sometimes enhanced by social
features like performance comparison. We see a demand
for a goal-driven support of fitness goal achievement to be
addressed by a digital coach. The digital coach identifies
strength and weaknesses of the subject, generates a training plan, motivates and helps, just like a real coach. Such
a digital coach will highly benefit from the activity tracking
system data which is used to personalize the training plan
based on performed activities.
Author Keywords
Pervasive Computing; Personal Informatics; Health; Personal Virtual Assistant.
ACM Classification Keywords
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than ACM
must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,
to post on servers or to redistribute to lists, requires prior specific permission and/or a
fee. Request permissions from [email protected].
UbiComp/ISWC ’15 Adjunct, September 7 - 11, 2015, Osaka, Japan.
Copyright 2015 © ACM 978-1-4503-3575-1/15/09...$15.00.
DOI: http://dx.doi.org/10.1145/2800835.2800961
H.5.m [Information interfaces and presentation (e.g., HCI)]:
Miscellaneous
Introduction
During the last years, activity tracking systems have become a usual accessory for athletes. The devices go together with the promise of an increased self-awareness to
be followed by a higher feasibility of physical fitness goals.
Thus, the devices address technical and motivational challenges. From a technical perspective, there are devices
which are moderately accurate in detecting simple activities such as steps while walking, however, more complex
derived information like calculated distance might be highly
erroneous [11]. Despite accurate data, studies have shown
that users frequently stop using respective devices [4]. We
assume that this lack of long term motivation is closely
linked to the difficult task of interpreting the data in actually
supporting a personal fitness goal.
Raw performance information should be connected to a
personal fitness goal and be used to help the individual understand what needs to be done to achieve the goal. Our
assumption builds on the existing work on the benefit of
goal setting and motivation: The conscious decision for a
goal and the motivation for activities involved in goal realization enable the goal realization process [9].
In a professional training setup a personal fitness coach is
responsible for goal setting, motivation and training. Studies
have shown that such a personal fitness coach has positive influence on the adherence and motivation of doing
training exercises [7, 10]. Following this idea, we envision a
personal digital coach to extend the features of an activity
tracking system by personal goal selection, personalized
and adaptive training plans and motivation. In the following,
we outline an early concept of a digital fitness coach.
To build such a system, it is important to consider the limitations of self organized training. When people choose a
training plan on their own, they often set high, unrealistic
fitness goals due to inaccurate self-estimations and high
motivation. People tend to execute the training exercises
initially but later the the probability of deviation from the
plan in terms of missed training units increases because
it is difficult to keep up with the schedule. At this point, a
personal trainer might adapt the schedule according to observations and the feedback he receives from the subject.
These adaptations might be in terms of changing the difficulty of the training units, recommending different units or
both.
Existing support on training plan creation1 allow individuals
to set an initial goal such as "run 5 km" and later set time
constraints e.g. run the same distance under 15 min or 35
min. However, this neither takes the actual physical performance into account nor considers preferred training units
(e.g. walking, running, climbing) or weather conditions.
The envisioned digital coach analyses the performance
data collected by the activity tracker, in order to check if the
training was executed properly. In case of deviations from
the training plan, the system will adapt the training plan to
the preferences of the individual – recommend different
training units for the next day or change the whole training
plan to a lower or higher level of difficulty.
A Digital Fitness Coach
Technology for Digital Coaches
The goal of our work is the extension of activity tracking
systems to personal digital coaches. The digital coach supports the process of goal identification, generates a training
plan for goal realization, adapts the plan to individual performance and motivates the subject.
To realize personal virtual assistant in domains like assisted
living Partially Observable Markov Decision Processes
(POMDP) and Markov Decision Processes (MDP) have
shown good performance [3]. A POMDP is defined by a tuple < S, A, T, C, I, O, µ >. S is a finite set of states, A is
1
See for example resources at [1].
a finite set of actions, T (s, a, ś) is a transition distribution,
C an action cost function, I is an initial state distribution, O
is a finite state of observations and µ(o|s) is a distribution
over observations o.
In the following, we outline the use of POMDPs and MDPs
for the adaptation of the training plan based on data from
an activity tracking system. We formulate the problem of
selecting the right training plan as an assistant POMDP. S
represents a set of training plans, A represents the actions
and can be keepT, chooseLighterT, chooseHeavierT .
keepT means, that the actual training plan is ideal for the
person and will not change it. In the training plan "run 5
km in 25 min", the action "chooseLighterT" will select an
easier training plan, for example "run 5 km in 30 min" and
"chooseHeavierT" will find a tougher training plan, for example "run 5 km in 20 min". In I the initial training plan is
stored, which will be suggested for a new user according to
his self-declared fitness level. In O the observations from
activity tracking system are stored which will indicate if the
current training is ideal, too light or too heavy. µ denotes
the observations of training unit executions in a training
plan.
The Markov Decision Processes (MDP) [2] are used to
model each training plan. A MDP is defined by a tuple
(SM DP , AM DP , PM DP , RM DP , γ)). SM DP is a set of
possible world states representing the progress in the training plan e.g. "Day-5" which represents the fifth action in the
plan. AM DP is a set of possible actions which can be for
example "no-activity", "run-less-than-5000-steps" or "runmore-than-5000-steps". The transition possibility PM DP
indicates the probability, that the persons will execute a
training unit at a certain state. For every state there is also
a reward function R(s, ś) which represents the reward after
transition to state ś from s. γ is the discount factor bounded
between 0 and 1 and indicates the importance of future rewards.
The solution of the MDP is to find an optimal policy π that
specifies the action π(s) that the Digital Fitness Coach will
choose in state s representing the best training unit advice
for the users.
A key feature of the approach is that by using POMDP for
training plan selection the system can reason over the observations and can approximate the right training plan for
the users. POMDPs have been successfully used in advanced planning and decision making approaches, e.g. for
helping people with dementia [5].
The MDP is helpful because it can be used to model training plans. It is also suitable to include the uncertainty of
transitioning to the next state when a training unit advice,
based on the person’s actions, is provided by the Digital Fitness Coach. The Digital Fitness Coach can learn the users’
behavior by observing user actions and then updating the
transition possibilities between states after training unit advices are provided.
We envision an implementation of the smart fitness coach
which consists of a multi-module approach presented in
1. The Sensordata Collector will fetch the wearable device data daily. The Movement Analyzer will enrich the
data with additional tags, based on the step counter data,
for recognizing different activity types. The Belief Monitor
analyzes, if the user has executed the advice provided by
the Digital Fitness Coach and decides if the running training
plan should be kept or an easier or more difficult training
plan should be chosen. In order to help users in achieving
their goal the Policy-module decides what the next advice
for a training unit is. Through the Communication to usermodule the next actions are displayed to the users.
Systems which integrate goals generally require the user to
individually enter information about the training [9, 7]. The
fine grained data collected by activity tracking systems is
not yet used for training plan adaptation or user motivation.
This data is mainly considered as a complementary data
type made accessible to the user [6].
Figure 1: Overview of the Digital Fitness Coach model
Based on the goal set by the user e.g. "run 5 km under 20
min", the Digital Fitness Coach will provide the first predetermined training unit. On day one, the user executed the
first training unit "walk-10000-steps". The system will move
to the next state in the training plan while the POMDP will
keep the current training plan for the user. The next advice
according to the training plan is "walk-12000-steps and run
20 min". At the end of the day the user walked just 8000
steps and did not run. The Digital Fitness Coach would
now update the transition probabilities for the advice "walk12000-steps and run 20 min" " and change the state in the
POMDP for a better suitable training plan which could be
an easier one: "run 5 km under 22 min". Important to mention is, that the user does not need to start the new training
plan from the beginning. He will be placed to the right training state, considering the effort he already invested and the
time between the last active day.
Related Work
Existing work on fitness goal externalization and training
motivation does not consider the actual performance of the
user.
One of the few examples for systems which adapt a training
plan based on user performance is Gymskill [8]. The device analyzes the human interaction with gym equipment,
analyzes human movement and provides suggestions for
the individual performance. The digital coach as extension
of activity tracking systems follows this idea with a focus on
the personalization of training plans and motivation.
Conclusion
In this paper, we have outlined a Digital Fitness Coach to
support individual fitness goal achievement based on training plan generation, adaptation and user motivation. With
this system we intend to overcome motivation problems
in training. The system uses the fitness goal data combined with performance data collected with activity tracking devices. Technical core of the approach is the use of
POMDPs and MDPs for the adaptation of training plans
based on the collected training data. Currently, we are
working on the first prototype of such a system. We plan
to build different POMDPs and MDPs to realize training plan
adaptation and will identify the respective performance in
comparison with a personal fitness trainer and an activitytracker-only setup.
Acknowledgement
This work has been funded by the LOEWE initiative (Hessen, Germany) within the NICER project
<https://www.nicer.network>.
REFERENCES
1. 2015. Runners World. (June 2015).
http://www.runnersworld.de/
2. Richard Bellman. 1957. A Markovian Decision Process.
Indiana Univ. Math. J. 6 (1957), 679–684. Issue 4.
3. Alan Fern, Sriraam Natarajan, Kshitij Judah, and
Prasad Tadepalli. 2014. A decision-theoretic model of
assistance. Journal of Artificial Intelligence Research
50 (2014), 71–104.
4. Daniel Harrison, Nadia Berthouze, Paul Marshall, and
Jon Bird. 2014. Tracking Physical Activity: problems
Related to Running Longitudinal Studies with
Commercial Devices. In Proceedings of the 2014 ACM
International Joint Conference on Pervasive and
Ubiquitous Computing: Adjunct Publication. 699–702.
5. Jesse Hoey, Pascal Poupart, Axel von Bertoldi, Tammy
Craig, Craig Boutilier, and Alex Mihailidis. 2010.
Automated handwashing assistance for persons with
dementia using video and a partially observable
Markov decision process. Computer Vision and Image
Understanding 114, 5 (2010), 503–519. DOI:
http://dx.doi.org/10.1016/j.cviu.2009.06.008
6. Randy Klaassen, Rieks op den Akker, Tine Lavrysen,
and Susan van Wissen. 2013. User preferences for
multi-device context-aware feedback in a digital
coaching system. Journal on Multimodal User
Interfaces 7, 3 (2013), 247–267. DOI:
http://dx.doi.org/10.1007/s12193-013-0125-0
7. Matthias Kranz, Andreas Möller, Nils Hammerla, Stefan
Diewald, Thomas Plötz, Patrick Olivier, and Luis
Roalter. 2013. The mobile fitness coach: Towards
individualized skill assessment using personalized
mobile devices. Pervasive and Mobile Computing 9, 2
(2013), 203–215. DOI:
http://dx.doi.org/10.1016/j.pmcj.2012.06.002
8. Andreas Moller, Luis Roalter, Stefan Diewald,
Johannes Scherr, Matthias Kranz, Nils Hammerla,
Patrick Olivier, and Thomas Plotz. 2012. GymSkill: A
personal trainer for physical exercises. 2012 IEEE
International Conference on Pervasive Computing and
Communications, PerCom 2012 March (2012),
213–220. DOI:
http://dx.doi.org/10.1109/PerCom.2012.6199869
9. Sean Munson and Sunny Consolvo. 2012. Exploring
Goal-setting, Rewards, Self-monitoring, and Sharing to
Motivate Physical Activity. International Conference on
Pervasive Computing Technologies for Healthcare
(2012), 25–32. DOI:
http://dx.doi.org/10.4108/icst.pervasivehealth.2012.248691
10. R M Ryan, C M Frederick, D Lepes, N Rubio, and K M
Sheldon. 1997. Intrinsic motivation and exercise
adherence. International Journal of Sport Psychology
28, 4 (1997), 335–354.
11. Judit Takacs, Courtney L. Pollock, Jerrad R. Guenther,
Mohammadreza Bahar, Christopher Napier, and
Michael a. Hunt. 2013. Validation of the Fitbit One
activity monitor device during treadmill walking. Journal
of Science and Medicine in Sport 17, 5 (2013),
496–500. DOI:
http://dx.doi.org/10.1016/j.jsams.2013.10.241