Paper: Entertainment modeling through physiology in physical play

Written by Milad Ghaznavi

A child playing Bug-Smasher


Yannakakis, Georgios N., and John Hallam. “Entertainment modeling through physiology in physical play.” International Journal of Human-Computer Studies 66.10 (2008): 741-755.


The goal of this paper was to provide an entertainment children-user model which predicts fun when children play a physical game. This entertainment model would predict the fun, by analyzing physiological measures.
They had two challenges to overcome for creating this model. The first challenge was how can we create this model, I mean using which existing technique? And the second challenge was how can we eliminate the impact of physical activities from captured physiological measures (I mean a physical activity and a physical game have some common features that are noises in prediction model)?
To producing this model, the authors designed two experiments, and used some feature selection algorithms and classification mechanisms. In the other words, they used and compared two feature extraction algorithms and three classification algorithms.

At the end, an entertainment model generates a number y that shows how much “fun” is it. For example in comparison of two games, the more fun game gets the higher value.

Main Experiment (the first experiment)

Seventy two normal-weighted children whose ages cover a range between 8 and 10 years participated in the main experiment. They played Bug-Smasher in nine variants on the Playware playground. The description of Bug-Smasher game is as follows:

The “Bug-Smasher” game is used as the test-bed game in the experiments presented here. Bug-Smasher is developed on a 6 x 6 square tile topology. During the game, different “bugs” (colored lights) appear on the game surface and disappear sequentially after a short period of time by turning a tile’s light on and off, respectively. A bug’s position is picked randomly according to a predefined level of bug spatial diversity. The child’s goal is to smash as many bugs as possible by stepping on the lighted tiles.

You can also see this playground in the Figure-1.

A child playing Bug-Smasher

Figure-1: A child playing Bug-Smasher

The children each time played a pair of games on the playground. The general phases of experimental setup is shown in the Figure-2. After completing each round of game, a child was asked “Which game was more fun?“.

Figure-2: General phases of the experimental setup followed

Figure-1: General phases of the experimental setup followed

During game play, the following physiological measures were captured:

  • Heart Rate (HR)
  • Blood Volume Pulse (BVP)
  • Skin Conductance (SC)

Controlled physical activity experiment (The second experiment)

This experiment was designed in order to control the physical engagement of the players. In a physical game context, the degree to which a game is entertaining influences the physical engagement of the player, and hence the intensity of the physical activity as well as (possibly) the kind of physical activity. They designed a non-entertaining variant of the Bug-Smasher game named the “Stomping game”. The Stomping game is as follows:

“Children are asked to stomp on a different one of four constantly lighted tiles of different color each time they hear a sound coming from the game platform. The four tiles are placed at the corners of a 3×3 square in the center of the 6×6 platform; two tiles equals the average distance between bugs appearing in the Bug-Smasher game. The sound determining the frequency of child-game interaction occurred at a rate equal to the average of the bug appearance rates of the two different levels of challenge used in the Bug-Smasher game.”

This game is completely predictable and interaction is absent in that the playground does not react to the child’s action.
For the control experimental protocol, we asked 18 naive normal-weighted children (nine boys and nine girls) aged 8–10 years to play five games each on the Playware platform. The set of five games played comprised four games of Bug-Smasher, in two pairs, and the physical activity control game referred to above. As in the main experiment, two game variants with differing levels of challenge and curiosity were played in both orders, giving four Bug-Smasher variant games plus the control game.
The same questionary and physiological measures as the main experiment were captured.

Entertainment Model

The authors created several models by eliciting some other features from physiological measures and applied two feature extraction algorithms for finding an appropriate set of features, and three different classification mechanisms in order to predict fun.
They used nBest and Sequential Forward Selection (SFS) algorithms for finding a suitable set of features that are better and more informative as an input for the prediction entertainment model. For the prediction model, they used three algorithms, Large Margin Algorithm (LMA), Meta-LMA and Artificial Neural Network(ANN).


The authors reported that children’s reported entertainment preferences correlated well with specific features of the recorded signal. Exactly, the features derived from HR and BVP correlates with the preferences.

By comparing these machine learning algorithms they reported that ANN combined with SFS had the best accuracy in predicting a child’s preferences, and the best feature set was {HF, E{h}, σ{RR}}. By using {HF, E{h}; σ{RR}} subset as an input and using SFS as feature selection algorithm and ANN as prediction model, they achieved to %79.76 accuracy. This model was abled to distinguish between an entertainment and a non-entertainment physical activity and also to predict children’s preferences.


This paper discussed a crazy work, which creates an entertainment model predicting the fun when children play a physical game. The idea was really amazing, and in my opinion had really crazy aim. In this part of my report, I want to discuss strengths and weaknesses of this paper.


As I mentioned before, in my opinion the idea of this paper was really amazing. Predicting an emotion using physiological measures is one of craziest topics that a researcher can work on. They had done a big work. Long time for capturing the data, lots of statistical works and using complicated algorithms and mechanisms to create a prediction model.
They recruited 72 children for the main experiment, and this number of participants helps to do more accurate study.
In doing experiment they actually considered every thing and all details. For instance, in their experiment design and choosing the experiment environment, they mentioned that:

“Day-dependence and methodological conditions in capturing and classifying emotions when using physiological signal data raised by Picard et al. (2001) are satisfied in the work described in this paper. All experiments described meet three of the five factors for eliciting genuine emotion in the most natural setup: the experiments took place in a setup close to the real-world since children played in their school classroom, our emphasis was on internal feelings and subjects were not aware of the purpose of the experiment (other-purpose). Note that the study presented here is not focussed on the investigation of the long-term realistic physiology of children with regards to entertainment but rather the construction of a predictor of reported entertainment based on individual physiological signal features.”

They also studied the impact of the order of presenting the tasks, and showed that the measures were not influenced by them. This level of details and consideration in designing the experiment and analyzing the data is really amazing.


Besides all good points, the paper has some negative points that I discuss them in this section.
In reading the paper, I had some difficulties in understating the sentences the authors used. The language they used was really hard to understand, and you need to read a sentence several times to understand what the authors want to say. You can find lots of sentences with the long lengths. Another problem was that the paper was not well-tructured and not easy to follow.
As a technical viewpoint, the paper has some ambiguities. As we know, there are lots of ANN systems, and they did not any explanation that which ANN they used.

Because the output of the model is a numeric value, they can compare how much fun are games, but the authors did not provide any explanation that how they classified the results into two classes, entertainment and non-entertainment. An ANN mechanism can provide a number as an output, but the authors do not provide any further information that how they mapped these values to the goal attributes, entertainment activity and non-entertainment activity.

Another issue with this paper is that, in the second experiment, they designed a physical activity named “The Stomping game” and say that because this is predictable and is not interactive, so this is not an entertainment activity. They also provide that because 22 out of 24 cases in the second experiment preferred the Bug-Smasher, the Stomping game is not entertaining. But we can criticize this statement that may children prefer the Bug-Smasher game, but it does not mean that the Stomping game is not an entertainment activity. I mean that the reports just show that children preferred the other game, it does not mean that they saying that the Stomping game is not entertaining. They used this fact for labeling the data and by this criticism of training data, may we can discuss about the accuracy of the evaluation method.


As I mentioned before, this paper is interesting and in explaining every thing was in details. However, the paper had some negative points, but at last we can say that the positive points do compensate them.

 Further reading

Yannakakis, Georgios N., John Hallam, and Henrik Hautop Lund. “Entertainment capture through heart rate activity in physical interactive playgrounds.” User Modeling and User-Adapted Interaction 18.1-2 (2008): 207-243.

Mitchell, Tom M. “Machine learning. 1997.” Burr Ridge, IL: McGraw Hill 45 (1997).


One thought on “Paper: Entertainment modeling through physiology in physical play

  1. Brittany Kondo

    The authors of this paper reported a 79% prediction accuracy of their entertainment model. I noticed there is a discrepancy between the prediction accuracy ratings. I briefly mentioned this in my review, where the authors of this paper described their reported accuracy (78%) of an anxiety-based model, to be “high”. However, both papers state that the acclaimed accuracy levels were supported in the literature. In my opinion, 79% seems low. In reference to your comment regarding the classification of games as entertaining and non-entertaining, I agree with you in that eliminating the entertainment factor in games is over-ambitious. An alternative classification could involve variations in level of entertainment (e.g., less or more entertaining).

Comments are closed.