Note for Lecture 'Human Robot Interaction' at Tokyo University of Agriculture and Technology.

Information Transmission and Impression Formulation

in Human-Robot Interaction

Lecturer: Toru Nakata (toru-nakata@aist.go.jp)

May 14. 2004.

Abstract

The author discusses and reports on the relationship between characteristics of reaction behavior and impression formulation in human-robot interaction. An information transmission efficiency index is employed to describe the characteristics of various responsive behavior of a robot. A high informational transmission efficiency signifies that the robot responds with a fixed reaction. A low informational transmission efficiency means the robot chooses responses randomly. In a society of animals, information transmission efficiency relates to social communication. Since information transmission efficiency is controllable by a behavior algorithm, robots can control the formulation of their social impressions by adjusting these efficiencies properly. In the experiment reported here, three kinds of randomness of response algorithms for a robot were prepared with informational transmission efficiencies of 100%, 50%, and 0%. In this experiment, a high information transmission efficiency gave the impression of the existence of intelligence. A playful impression regarding the robot was produced the most in interactions using the intermediate transmission efficiency.

(WMV Format, 1 MB) Interaction with Low Information Transfer Efficiency	(WMV Format, 1 MB) Intermediate Efficiency
(WMV Format, 1 MB) High Efficiency	(WMV Format, 0.8 MB) Reverse Information Transfer
(WMV Format, 1 MB) Human Robot Interaction Experiment

1. Introduction

Many researchers and companies have been trying to build robots and artificial beings that interact with humans (Bates, 1994; Ogata and Sugano, 2000). Interactions with pet robots and characters in video games represent a new style of entertainment.

These robots and characters are expected to communicate with humans in a natural way (i.e. communication through verbal and nonverbal way without keyboards or other artificial input devices). They are required not only to obey the user’s commands, but should also formulate lifelike impressions and attain a familiarity with users.

However, a methodology for the management of social psychological expressions by interactive artifacts has not been sufficiently developed. Many owners tire of their pet robots, even though their body actions are lifelike to some extent. They lack the ability to adjust their expressions adequately for social relationships and contexts of interaction.

The objective of this paper is to present a feasible method for robots to better control social psychological impressions when interacting with humans.

Readers may doubt that social psychological factors are quantitatively measurable. Certainly, highly complex factors such as jealousy are very difficult to measure.

Some low-level factors, however, have been studied quantitatively. The promptness of a response expresses familiarity, royalty and/or initiative (Gilbert and Hailman, 1966; Kameyama, 2000). The temporal length of touching relates an affinity between an infant monkey and its mother (Harlow and Zimmermann, 1959; Hinde and Atkinson, 1970). The spatial closeness of red deers correlates with their kinship (Clutton-Brock et al., 1982). Variables like time-length and spatial distance are simple and easily observed.

Furthermore, animal researchers have gathered other data for social communication analyses.

The informational index of interaction is one of the most utilized parameters for communication analyses (Losey, 1978; Lehner, 1996). Using the informational analysis method, researchers can measure flows of information in animal groups that relate to social structure, as shown later.

Robots that interact with humans should manage their behavior to produce the proper social impression. One of the solutions to this problem would be attaining proper control of informational factors when interacting.

This paper discusses the role of informational factors in human-robot interactions. After a brief explanation of informational analysis, the relationships between informational factors and social psychological factors are discussed with results of previous studies on animal communication. This study also aims to expand the informational phenomena of animal communication to human-robot-interactions via experiment.

2. Information transmission in social communication

2.1 Definitions of informational indexes

The word information usually means knowledge or data. In informatics, information is defined as the level of uncertainty reduction of probability variables.

Behavioral information can be defined as uncertainty reduction in action sequences of animal communication. In general, the behavior of an individual varies, and future behavior is uncertain. In social interactions, the behavior of an individual often is influenced by an opponent’s behavior. By knowing the opponent’s previous behavior, one can make a better prediction of future behavior.

Behavioral information is mathematically defined as the following. Suppose 2 individuals are interacting with each other (Fig. 1). We try to measure the information transmission from the initiator to the responder.

Fig. 1. Dyad interaction schema.

Assume the initiator can perform several kinds of actions, which are defined as. Likewise, the behavior categories of the responder are defined as.

The notation expresses a dyadic interaction that is a sequence of and actions. The frequency of is written as. The matrix consisting ofis called a sociometric matrix in the behavioral field.

Table 1 shows a pseudo example of a sociometric matrix for explanation purposes. Using the notation above, stands for the frequency of the dyad that the responder fluttered its legs after the indicator shook the responder’s body.

Table 1. A pseudo example of a sociometric matrix.

Informational indexes are calculated by the following formulas (Lehner, 1996):

…. (1)

…. (2)

…. (3)

…. (4)

… (5)

… (6)

… (7)

… (8)

… (9)

In ethological terminology, is called the behavioral uncertainty of the initiator, and is the behavioral uncertainty of the responder. is the amount of transmitted information from the initiator’s action choice to the responder’s reaction choice.

D and C represent normalized transmitted information and are called information transmission efficiencies. Information transmission efficiencies are the ratio of the influence of the initiator to the responder’s choice of reaction. If the responder neglects the initiator, the information transmission efficiencies are nearly 0. Conversely the reaction selection depending on the initiator’s action enlarges the information transmission efficiencies.

The information transmission efficiency D represents the percentage of how often the responder recognizes the initiator’s action. In this paper, D is called the distinguishablility of the responder. In the example, the distinguishablility is 49%. That means the responder responds to 49% of the initiator’s action.

C signifies the proportion of how often the initiator can direct the responder’s reaction. C is called the controllability of the responder. The controllability is 54% in the example, meaning the initiator can select and initiate the responder’s reaction with 54% probability.

2.2 Behavior of Information transmission in animal communications

Informational analysis of social communication was frequently used in animal studies during the 1970’s (Table 2).

Table 2. Information transmission of animal communication.

The researchers found that information transmission increases as the interaction proceeds (phase I in Figure 2) in inter-male interactions of crabs (Rubenstein and Hazlett, 1974; Hazlett and Estabrook, 1974), grasshoppers (Steinberg and Conant, 1974), and shrimps (Dingle, 1969). For example, a threatening action of an individual evokes a counter-threat from the opponent. This threat-to-threat loop becomes stead, and amount of information transmission will be enlarged.

The loop is dissolved after dominance is determined, and information transmission then diminishes (phase II in Fig. 2). But the winner maintains a retaliatory policy against any aggressive display by the loser. Therefore, information transmission from the loser’s behavior choice to the winner’s reaction choice remains somewhat high (Hazlett, 1980).

Thus, information transmission signifies the existence of a standing policy or awareness towards other individuals.

This phenomenon is also observed in the behavior of human infants. According to Watson (1977), human infants select their reactions towards unknown machines based on the behavior of the machines. This tendency enlarges the information transmission from the behavior of unknown machines to a known behavior choice. After human infants lose interest, the information transmission diminishes.

Fig. 2. Typical growth pattern of information transmission in interaction progress.

2.3 Robot and informational transmission

Information transmission is determined by the frequencies of the dyads. It would be easy for robots to control information transmission by adjusting the frequencies of dyads.

In addition, information transmission relates to social meanings such as dominance and awareness. This generalization seems valid even in human-robot interactions.

Properly controlled information transmission would produce good social impressions. This expectation is worth being examined experimentally.

3. Experiment

3.1 Purpose

The purpose of this experiment is to examine relationships between the informational characteristics of a robot and the social impressions produced when interacting with humans.

The following social impressions were selected as targets of the analysis.

1. Impression of existence of intelligence in the robot.

2. Playfulness of the robot.

3. Cuteness of the robot.

These impressions are typical kinds of instantaneous social impressions. These shallow social impressions can be analyzed, even though the experiment consists of short sessions without any particular context.

3.2 Hardware of the experimental robot

3.2.1 Appearance

The employed robot has the appearance of a stuffed teddy bear (Fig. 3). The reason of selecting this appearance is that many of recent robots or interactive toys have been made with stuffed animal bodies. The result of this experiment will be useful for those interactive stuffed-bodied toys. Also, the author does not have any practical reason to dare to employ other uncanny appearance as the robot’s appearance.

Appearance can strongly affect impression formulation. To eliminate any appearance effects, the procedure of evaluating impression should be constructed as described later.

Fig. 3. Experimental Robot.

3.2.2 Actuator

The four limbs of the robot can be bent and stretched automatically. To give a lifelike impression, the movements of the limbs should be lifelike also. Each actuator has two degrees of freedom in the rotations, and the distance between the axes of rotation is 15 mm. This small gap makes the movement like a 2-directional pin-jointed rotation that is similar to the movements of the shoulder and hip joints animals.

Fig. 4. Sensors and actuators of the robot.

Fig. 7. Actuation structure of the robot.

3.2.3 Sensors

3 kinds of sensors are installed in the robot:

• A sound sensor is installed at the ear position of the robot in order to detect the voice of a subject. This equipment is a set of PS-3001S sound sensors from EK Japan Inc.

• An acceleration sensor is installed in the body to detect shaking. ADXL202 sensors of Analog Devices Inc are used.

• A touch sensor is installed in the forehead of the robot to detect touch. The area of sensing is a trapezoid, as shown in Fig. 5. The sensor is made of a conductive sponge (F-10B of Hozan Inc.) of which the electrical resistance greatly changes according to the loaded pressure. .

Fig. 5. Area of the touch sensor installed in the robot’s forehead.

3.3 Software for Sensing and Actuation

3.3.1 Sensing human action

The robot samples data from the sensors every 0.1 seconds. The detection procedure is executed with the priority order shown in Fig. 6.

The following 4 categories of subject behavior were prepared:

1. Calling (A1). When the sound sensor detects a loud sound, the robot judges the sound to be the call of the subject. The threshold of loudness was adjusted for each session to prevent catching background noise.

2. Touching (A2). When a call is not detected, the robot monitors data from the touch sensor. When pressure over 10 kiloPascals is detected at any position in the touch sensor area, the robot judges the subject to be touching the robot’s forehead.

3. Shaking Body (A3). When a subject’s touch is not detected, the robot monitors the acceleration sensor. If the robot body is rapidly moved, over 40 , the robot judges the subject to be shaking the robot’s body.

4. Silence (A4). When no behavior is detected for 4.0 seconds, the robot judges the subject to not be interacting with it. The robot makes no reaction before 4 seconds pass.

Fig. 6. Sensing algorithm of the robot.

3.3.2 Robot’s body movement patterns

The following four patterns of body movements were designed as the robot’s output towards a subject:

1. Raising arms (B1): The robot raises both arms together and holds the posture for 1.0 second (Fig. 9).

2. Fluttering legs (B2): The robot flutters its legs twice alternately for 1.0 second (Fig. 10).

3. Curling (B3): The robot curls all limbs inside and holds the posture for 1.0 second.

4. No reaction (B4): The robot limits power to all actuators and frees all limbs for 1.0 second.

To avoid any of the body movement patterns from implying any particular meaning, the body movement patterns were short and rather meaningless.

3.3.3 Reaction selecting algorithms

As defined above, the behavior categories of a subject and the robot were defined by the following:

…. (10)

… (11)

The response algorithm of the robot was designed as follows.

Prior to any action by the subject, the robot does not move and all limbs are free.

After detecting an action by the subject, the robot stochastically selects a reaction.

The following 3 algorithms for reaction selection by the robot were prepared.

1. Non-random reaction algorithm: A certain reaction is non-randomly selected against each type of the subject’s actions (Table 3).

2. Semi-random reaction algorithm: Two reactions each have a 50% probability of being selected based on the action of a subject (Table 4).

3. Full-random reaction algorithm: The robot’s reaction does not depend on the action of the subject. Selection probabilities of all reactions are all set to 25% for all actions (Table 5).

Table 3. Response probabilities under the non-random reaction algorithm. ()

Table 4. Response probabilities under semi-random reaction algorithm. ()

Table 5. Response probabilities under full-random reaction algorithm. ()

The characteristics of the reaction algorithms can be described with information transmission efficiencies.

The actual values of information transmission efficiencies will differ among subjects and are insufficient for signifying informational features of the reaction algorithms..

Instead of actual values, maximum likelihood estimators of the information transmission efficiencies were used. The maximum likelihood estimators of the information transmission efficiencies can be regarded as pre-experimental expectations of the efficiencies. In general, an expectation of the information transmission efficiency depends on the distribution of occurrence probabilities of the preceding actions. However, thanks to symmetries in the selection probability matrices, all expectations of the 3 reaction algorithms were determined without dependence on the subject’s action selection in this experiment.

The maximum likelihood estimators of distinguishablility and controllability are represented by and .

The non-random reaction algorithm sets , the semi-random reaction algorithm sets , and the full-random reaction algorithm sets .

3.4 Experimental Procedure

Each experimental session was conducted with 1 subject individually.

An experiment proceeds according to the following steps:

Step 1) The subject is educated about the sensors of the robot as follows:

• The robot can be stimulated by calling, touching the forehead, or shaking the robot’s body. Also, the robot may move without any cue.

• The robot does not distinguish the intensity or direction of the subject’s actions.

• The robot reacts following an action by the subject. Sometimes the subject’s actions may be ignored.

Step 2) The subject is educated about the nature of the experiment:

• The subject will have 3 sessions.

• In each session, the subject will be instructed to hold the robot and play with it (Fig.8).

• The behavior of the robot may differ from session by session.

• Each session lasts 90 seconds. (Long session may result in fatigue. The author expects a dyad to take about 3 seconds. So, 30 dyads are expected to occur every 90 seconds. That is enough for impression formulation and correcting the amount of data for informational analysis.)

Step 3) The sessions starts. The robot behaves according to one of the 3 response behaviors in each session. The order of the response behaviors was randomized over the subjects to cancel any order effect.

Step 4) The subject evaluates the intensity of the following 3 impressions of the robot: impression of existence of intelligence, playfulness, and cuteness of the robot.

The evaluation method is based on relative-scoring that cancels offsets of impressions made by static features of the robot. The subject ranks the intensities of the impressions as ‘strongest’, ‘intermediate’ or ‘weakest’ among the 3 types of response behaviors.

Fig. 8. Scenes of the experiment.

Fig. 9: Reaction of raising arms.

Fig. 10: Reaction of fluttering legs.

Fig. 11: Reaction of curing.

Fig. 12: Reaction of release and rest.

3.5 Results

3.5.1 Constituents of the subjects

The group of subjects consisted of 20 Japanese with the demographics shown in Table 6.

It is generally assumed that 20 subjects are enough to observe any major psychological tendencies in the results.

Table 6. Constituents of the Subjects

Job	Female	Male	Sum
Students	2	8	10
Office worker	5	5	10
Sum	7	13	20

3.5.2 Informational Result

Table 7 shows the averages and standard deviations of the informational results over all subjects.

The numbers of dyads in each session was usually over 50. This is enough data for analysis. The average duration of a dyad was 1.7 seconds.

The observed informational transmission efficiencies were very similar to the maximum likelihood estimators calculated prior to the experiment.

The information transmission efficiencies from an action choice by the robot to the following action of a subject were less than 0.2. This means that less than 20% of the subjects’ actions depended on the preceding robot action.

Table 7. Means and standard deviations of informatical indexes of the observed data.

3.5.3 Impression of intelligence

Figure 13 shows the results of the impression intensity ratings of intelligence existence in the robot.

The non-random reaction algorithm gained 11 votes in the strongest intensity impression category. The difference between the answer distributions of the non-random algorithm and the semi-random algorithm was verified to be statistically significant by using Mann-Whitney’s U-test at a significance level of 0.1.

Fig. 13. Information transmission efficiency versus intensity of the impression of intelligence existence in the robot.

3.5.4 Intensity of playfulness

Figure 14 shows the results of the answer distributions regarding the playfulness impression.

Even though the non-random reaction algorithm got the fewest answers in the most-playful impression, this difference was not significant.

Fig. 14. Information transmission efficiency versus intensity of playfulness impression.

3.5.5 Intensity of cuteness

Figure 15 shows the results of the answer distributions regarding the cuteness impression.

On this questionnaire, answers were split into the 2 extremes of information transmission efficiency. The non-random reaction algorithm and the full-random algorithm got the ‘most’ and ‘least’ votes. The semi-random reaction algorithm got intermediate answers,

The significances of these tendencies were not verified statistically.

Fig. 15. Information transmission efficiency versus intensity of cuteness impression.

3.6 Discussion

3.6.1 Did the robot control the subjects?

The information transmissions in the robot-to-human direction were constantly low.

This means that the subjects always maintained the role of initiator in the interaction.

It is also observed that each dyad was rather independent to the previous dyad. Any long-sequence context was very weak in this experiment, and analyzing the interaction with short dyadic scope was enough.

3.6.2 Impression of intelligence

The intensity of the impression of intelligence existence had a positive correlation with the informational transmission efficiency. This relationship translates into the expression of intelligence or awareness in robots.

3.6.3 Impression of playfulness

The non-random algorithm tended to produce weaker playfulness impressions than the semi-random and the full-random reaction algorithms.

The semi-random algorithm had slightly better results than the full-random algorithm. A similar phenomenon is report by Bates (1994). When a character displayed on a computer screen behaved unexpectedly due to faults in its software, the human audience felt that the character was unique and interesting. In contrast, other characters behaving under deterministic control were not seen as attractive.

Uncertainty in behavior may have the power of attracting human’s attention. However, the results in this experiment were not statistically significant.

3.6.4 Impression of cuteness

The subjects’ feelings on this impression were greatly split.

The full-random reaction algorithm may have produced a mischievous impression of the robot. While 4 subjects out of the 5 female office workers felt the full-random was the cutest, a significant relationship between the preference and property of the subjects was not found.

It can be said that the informational transmission efficiency of the robot towards humans may affect feelings of cuteness or other familiarity-related impressions. Ideal robots that interact with humans should adjust their informational transmission efficiency to the value that the human prefers.

3.6.5 Effects by static factors in the experiment

This experiment was carried without particular context. Emptiness of context may affect impressions.

Usually we can expect that effects of static factors are canceled by relative evaluation of impressions. However, static factors of robots sometimes produce uncanny impression.

The author thinks that existence of uncanny impression disturbs the intensity order of other impressions like uncanny valley phenomenon (Mori, 1982) and makes relative evaluation invalid.

We should be careful about what the subjects felt or thought about the robot. The experimenter recorded comments that the subjects made voluntarily. One subject felt strangeness since the robot’s behavior of any of the reaction policies is greatly different from real-life baby. Other 3 subjects felt mismatching between animal-like appearance and non-random reaction policy, since non-randomness made machine-like impression on the robot. These answers can be regarded as result of uncanny valley phenomenon; the subjects were confused, because expectations made by the robot’s appearance were greatly different with actual behavior of the robot.

The rest 16 subjects did not make comments relating to uncanny impression. For the majority of the subject, the appearance did not make bad impression effect. The author thinks the reason as the following.

What the subjects had perceived before the robot moves are teddy-bear appearance and soft touch feeling of the robot.

According to Lorenz (1943), appearance of round head with relatively-large eyes located lower position evokes watcher’s instinct that makes the watcher regard the object as an infant and be kind to it.

Harlow and Zimmermann (1959) reported monkey babies that were separated from their mothers tended to hug soft objects as substitute of mother. Soft touch may indicate familiarity.

The appearance of the experimental robot might make the subjects regard the robot as a infant or protection-requiring existence. Also soft touch of the robot might have produced familiarity to some extent. The author thinks that those static features of the robot had certain efficacy to suppress of uncanny impression.

4. Conclusion

Relationships between information transmission and production of social impressions in human-robot interaction were tested and discussed. A high information transmission efficiency formulated the impression of existence of intelligence in the robot. An intermediate information transmission efficiency produced a rather strong playful impression. The perception of cuteness of the robot depended on the subject.

The informational transmission efficiency of human-robot interactions is controllable by the software of the robot. It is a goal to design robots that produce more familiar impressions to humans by adjusting their informational interaction characteristics.

In the experiments reported here, there was no specific context. In future work, the synergic effects of contexts and information transmission should be examined, since robots are expected to work in the context-rich environments of human society having practical roles.

Appendix: Problems on measuring behavior

Informational analysis requires the proper observation of behavior. The incorrect categorization of behavior directly leads to erroneous results. The following considerations were made for the experiments in this paper. General advice was taken from the textbooks of Losey (1978); Martin and Bateson (1990); and Lehner (1996).

1) Length of scope of action sequence

An interaction is a sequence of actions. Social interactions sometimes have a long-spanning hysteresis. The selection of an action may be influenced by much earlier events. If so, this means the responder has a good memory and possess a sophisticated intelligence regarding action planning. Such a phenomenon is worthy of reporting.

However, the analysis of long action sequences over a dyad is very difficult because of shortage of data size (Fagan, 1977). Requirement on number of data for long sequence analysis swells exponentially as the order of interaction scope increases.

2) Temporal length of action

The time length of actions may differ from one another. Researchers always face the problem of whether double-length actions should be counted as two or not.

Time length can be ignored in some cases. Dawkins and Dawkins (1973) succeeded in finding information transmission in the behavior of chicks (Dawkins and Dawkins, 1973) and flies (Dawkins and Dawkins, 1976). Fentress (1973) also got meaningful results from the observation of mouse behavior.

In the experiment presented here, the robot reacts immediately after a human cue. The reactions of the robot interrupted the human action, so the lengths of the human actions were somewhat equalized.

References

• Bates, J., 1994. The role of emotion in believable agents. Communications of the ACM, 122-125.

• Clutton-Brock, T. H., Guinness, F. E. and Albon, S. D., 1982. Red deer behavior and ecology of two sexes, University of Chicago Press.

• Dawkins, R. and Dawkins, M., 1973. Decisions and the uncertainty of behaviour. Behaviour, 45, 83-103.

• Dawkins, R. and Dawkins, M., 1976. Hierarchical organization and postural facilitation: rules for grooming in flies. Animal Behaviour, 24, 739-755.

• Dingle, H., 1969. A statistical and information analysis of aggressive communication in the mantis shrimp gonodactylus bredini manning. Animal Behaviour, 17, 561-575.

• Fagen, R. M., and Goldman, R. N., 1977. Behavioural catalogue analysis methods. Animal Behaviour, 25, 261-274.

• Fentress, J. C., and Stilwell, F. P., 1973. Grammar of a movement sequence in inbred mice. Nature, 244, 52-53.

• Gilbert, B. K., and Hailman, J. P., 1966. Uncertainty of leadership-rank in fallow dear. Nature, 209(5027), 1041-1042.

• Harlow, H. F., and Zimmermann, R. R, 1959. Affectional responses in the infant monkey. Science, 130, 421-32.

• Hazlett, B., 1980. Patterns of information flow in the hermit crab calcinus tibicen. Animal Behaviour, 28, 1024-1032.

• Hazlett, B.A., and Bossert, W. H., 1965. A statistical analysis of the aggressive communications systems of some hermit Crabs. Animal Behaviour, 13, 357-373.

• Hazlett, B. A., and Estabrook, G. F., 1974. Examination of agonistic behavior by character analysis I: The spider crab microphrys bicornutus. Behaviour, 48, 131-144.

• Hinde, R. A., and Atkinson. S., 1970. Assessing the role of social partners in maintaining mutual proximity, as exemplified by mother-infant relations in rhesus monkeys, Animal Behaviour, 18, 169-176.

• Kameyama, M., 2000. Mental Process of guessing pet’s mind. Proceedings of the 6th annual conference of Human-Animal Relationship Society, 30. (In Japanese).

• Lehner, P. N., 1996. Handbook of ethological methods. Cambridge University Press.

• Lorenz, K., 1943. Die angeborenen Formen moglicher Erfahrung, Zs. Tierpsychol., Vol.5, pp.235--409, 1943. (in Tinbergen, N., 1951. The Study of Instinct, Clarendon Press, 1951.)

• Losey, G. S., 1978. Information theory and communication, in: Colgan, P. W. (Ed.), Quantitative Ethology. Wiley-Interscience.

• Martin, P., and Bateson, P., 1990. Measuring behavior. Cambridge University Press.

• Mori, M., 1982. The Buddha in the Robot. Charles E. Tuttle Co.

• Ogata, T., Sugano, S., 2000. Emotional Communication between Humans and the Autonomous Robot WAMOEBA-2 (Waseda Amoeba) Which has the Emotion Model. JSME International Journal, Series C: Mechanical Systems Machine Elements and Manufacturing, 43(3), 568-574.

• Rubenstein, D. I, and Hazlett, B. A., 1974. Examination of the agonistic behaviour of the crayfish orconectes virilis by character analysis. Behaviour, 50, 193-216.

• Steinberg, J. B., and Conant, R. C., 1974. An Informational analysis of the inter-male behaviour of the grasshopper chortophaga viridifasciata. Animal Behavior, 22, 617-627.

• Watson, J. S., 1977. The perception of contingency as a determinant of social responsiveness. Origins of the Infant's Social Responsiveness, Halsted Press Division of Wiley.

The End