Types of Communication

Communication may occur when a sender initiates in some channel a physical disturbance (the signal) that is detected by a receiver (fig 2-1). For example, a female firefly (sender) emits a pulse of light (signal in the optical channel) that is seen by a male firefly (receiver). Actually, mere detection by the receiver is not sufficient for communication, and later in this chapter a comprehensive and precise definition of communication will evolve. One might classify types of communication according to the communicants (sender and receiver), according to the channel, or according to other characteristics of the process. This volume concerns a limited group of communicational types, delimited as follows.

Fig 2-1. A simple communieational system, in which the communicants are represented as boxes connected by a channel represented as a line. Communication occurs when information is transferred via a signal in the channel from one communicant (the sender) to the other (receiver) as indicated by the arrowhead.

communicants

Senders and receivers may be animals, plants or even non-living entities such as computers. A classification based on the kind of communicants may be represented by a Venn diagram (fig 2-2) in which the universal set is that of all communicants. The major subsets are living senders and receivers, and their included subsets of animals. Each logical space in the diagram represents a kind of communication system, and the spaces in fig 2-2 are numbered for convenience. Logical space (1) is abiological communication, in which neither communicant is alive, as in signaling between two computers. The remaining spaces are types ofbiological communication, in which at least one of the communicants is alive. Included within biological communication are spaces (5) through (9), zoological communication, in which at least one of the communicants is an animal. The focus of this volume is logical space (9), in which both communicants are animals: animal communication.

Fig 2-2. Types of communicants may be used to classify communication. Each logical space in the diagram represents a different type of communication, numbered for reference in the text. Space (9) is the focus of this volume.

That animals may communicate with nonliving communicants and with plants is readily demonstrated by example. A star (sender) may send light (signal) detected by a noeturnally migrating songbird (receiver), which utilizes the signal to adjust its direction of flight (logical space 6) Both Smith (1968: 45) and Marler (1968: 103) discuss communication between plants and animals (spaces 7 and 8), as when the flower's nectar guides direct a bee to nectar (space 8).

channels

An obvious basis for classifying communication is by the physical channels through which signals are sent. To understand animal communication it is necessary to study all the channels simultaneously: optical, acoustical, chemical, tactual, etc. There appears to be no thorough and general study of animal communication in any channel, much less all the channels taken together, so it is useful to partition the general analysis of communication by channel. Although the focus of this volume is the optical channel, the strategy of analysis may provide a model for similar analyses in other channels.

animal communication

Animal communication as defined by fig 2-2 includes a broad range of phenomena, which may be further subdivided according to three characteristics emphasized in ethologicai literature. First, many authors (e.g., Wilson, 1968: 98; Tavolga, 1968: 273) draw attention to communication between different species, as in symbiosis, parasitism, predation, etc. This volume concerns primarily communicants of the same species: conspecific communication.

Second, Marler (1968: 103) and others have stressed the importance of "two-way" exchange of signals between the communicants, a notion I call communicative reciprocity. Not all conspecific communication is reciprocal in this sense; for example, one animal of a group may give an alarm call to which others respond only by fleeing. T. Johnston (pers. comm.) points out that linguists apply a notion of reciprocity to signals: if two individuals utilize the same set of signals or vocabulary of elemental units, these units possess a reciprocal quality. That reciprocity is a special case of communicative reciprocity in general, a case that may apply to certain kinds of animal signaling (e.g., aggressive threat signals of two individuals) but no to others (e.g., differences in signals used by male and female in courtship communication).

Finally, many ethologists emphasize the role of natural selection in structuring signals to maximize their efficiency and reliability in transferring information (e.g., Klopfer and Hatch, 1968: 31-32; Smith, 1968: 44-45; Marler, 1968: 103). Julian Huxley (1923) referred to highly evolved signals as "ritualized," a term ethologists apply to any behavioral patterns whose characteristics have been evolutionarily influenced because the patterns serve as signals. Although difficult to identify in practice, such ritualized communication has been a major focus of ethology.

Fig 2-3. Three properties of communication (circles) define various types of animal communication (see space 9 in fig 2-2). The logical spaces are denoted by letters for reference in the text, and the focus of this volume is space (h).

Figure 2-3 shows these characteristics of conspecificity, reciprocity and ritualization in a Venn diagram that defines eight logical spaces, all of which are subsets of space (9) in fig 2-2. Only in certain cases will it be easy to decide whether or not some instance of communication belongs unequivocally to a particular set, so these should be regarded as "fuzzy sets" presented only for the purpose of pointing out the principal kind of communication treated in this book. Space (a) might include the example of a hidden songbird watching a hawk soar overhead, (b) such a songbird watching one of its conspecifics moving about the environment, (c) the hawk chasing the songbird, and (d) a group of such songbirds maneuvering in a coordinated flock during flight. The remaining spaces involve more highly ritualized communication, and an example of (e) might be the display of eye-spots on the wings of a moth confronted with a predatory songbird. Space (f) includes the songbird giving ritualized threat display to a conspecific that shows no response, and space (g) includes the same bird threatening a member of another species in a mixed-species foraging flock, when the other bird responds in some way to the threat. Space (h), the mutual intersection of the three defining sets, I call social communication: reciprocal, ritualized communication between conspecifics. Social communication is the primary focus of this volume, and is roughly equivalent to Marler’s (1968: 103) "true communication" and Smith's (1968: 45) "social signal function."

Only certain groups of animals combine sufficient complexity of social behavior with keen enough vision to render their communication in the optical channel interesting and challenging to ethologists. This volume refleets the literature on those groups: insects, some other arthropods (such as spiders and crabs), cephalopod molluscs and most classes of vertebrates (bony fishes, amphibians, reptiles, mammals and birds).

signals

Charles Sanders Peirce, founder of semiotics, conceived of communication as taking place whenever a sign stands for something (its object or designatum) to somebody (the interpreter) in some manner (the interpretant or interpretation of the sign). C.W. Morris (e.g., 1946) translated these general notions into the physical world, in which a sign-vehicle stands for a referent and produces a response in the interpreter. Marler (1961), W.J. Smith (e.g., 1968) and other ethologists have used Morris's framework for analyzing animal communication, and recently there have been attempts to clarify and extend such uses (e.g., Stephenson, 1973; Johnston, 1976). In a simple example, the black mustache mark of the adult male common flicker ineastern North America (see figure 10 in Hailman, 1977a) stands for "adult maleness" (the referent) to other flickers (the interpreters) and produces observable effects on those interpreters. When Noble (1936) painted a mustache mark on a female flicker, her mate aggressively drove her away until Noble recaptured her and removed the mark.

Semiotics distinguishes between the signal, which is a collection of physical disturbances, and the included sign-vehicle, which is that aspect of the signal producing the response. For example, the Carolina anole is a small lizard in which the male extends his red throat-skin, called the dewlap, and makescertain display movements. Females watching a displaying male through glass partitions undergo ovarian development: a response to the optical signal. Crews (1975) found that surgical prevention of the movement abolished the response whereas changing the dewlap's color to black did not affect the response. The color is therefore not part of the sign-vehicle that affects female ovarian response. Ordinarily, ethologists do not know what part of an optical signal is producing the response; experimentation by altering the signal or by using models is required (see citations in Hailman, 1977a). It is also possible that different aspects of the same signal are different sign-vehicles, but this problem has been littie-studied in optical signals of animals.

In this book, I refer to the male flicker's mustache, the male anole's dewlap and other markings and structures--as well as certain movements, postures and orientations--as optical signals, even though such usage entails at least three problems. First, it blurs the useful distinction between the signal and its included sign-vehicle, but since the distinction cannot be made except in a few well-studied examples, there is little choice but to gloss over it with a call for more experimental studies. Second, the must ache mark or dewlap are not transmitted from sender to receiver: the light reflected from them is transmitted. Therefore, the markings, structures, movements and so on might usefully be called "signal-objects," but in general I shall make the distinction explicit only when confusion might occur. Finally, it is not always clear how one signal can be separated objectively from others in space or time, and there are cases in which one cannot specify "a" signal. Usually, this problem causes little trouble and may be made explicit when it does.

uncertainty

Information is the reduction of uncertainty, so one must first make precise the measurement of uncertainty. Players of the parlor game "twenty questions" quickly learn the optimum strategy of eliminating half the possibilities remaining with each question asked, which must be answered yes or no. If an array about which one is uncertain consists of the four members A, Β, C and D, the first question might be: is it either A or Β? If the answer be affirmative, one may ask if it is A; if negative, whether it is C. In either case, the correct item will be identified in two questions. An array of two items requires but a single question; an array of eight, three questions; an array of 16, four; and in general an array of n items requires Η questions when

which may be rearranged in logarithmic form as

The quantity H is the uncertainty (or entropy) of the array of n equiprobable items, and the units are called bits (binary digits).

It is easy to show that the binary strategy of guessing is the most efficient way to play the parlor game. For example, if the first question were "is it A?” the guess would be correct only in one-quarter of the games. If negative, one might then ask "is it Β?" which guess would also be correct in only one-quarter of the games. The third question "is it c?" will always remove all doubt in the remaining half of the games. The average number of questions required when using this serial strategy is therefore 1(1/4) + 2(1/4) + 3(1/2) =2.25 questions/game, whereas the “halving” strategy explained at the outset always requires exactly two questions. The efficiency of the binary technique helps one to feel intuitively that eq (2.1) is a useful expression for measuring uncertainty in arrays of equiprobable items.

Equation (2.1) gives a sensible answer only when the array is defined. If the task is to identify a particular playing card drawn from an ordinary deck without jokers, Η = log₂52 = 5.7 bits/draw. If the task is to identify only the suit of the card drawn, the uncertainty is less because there are only four equally probable suits: Η = log₂4 = 2 bits/draw. If someone removes certain cards from the deck before making the draw, the array of alternatives cannot be specified unless one knows what cards were removed; it may be impossible to calculate uncertainty. In most cases of animal communication, one cannot define the array and therefore cannot calculate uncertainty.

Suppose the items of the array are not equally probable, as might happen when someone removes cards from a deck. Even if one knows which cards remain, eq (2.1) can be employed only if the distinguishable items of interest (say suit) occur with equal frequency. In order to generalize eq (2.1), consider the fact that each equiprobable item contributes exactly 1/n portion of the total entropy, so that the equation may be rewritten

Since the probability of the ith item in the array is p_i = 1/n, it is also true that η = 1/p_i for the n equiprobable items. Substituting,

from which the fraction may be removed by noting that log 1/x = -log x, so

which is Shannon's general equation for entropy (Shannon and Weaver, 1949), often written with the negation sign removed outside the summation operator. One must accept intuitively that eq (2.2) delivers useful answers when the component p_i's are not all 1/n; the only requirement is that Σ p_i = 1.

In order to measure uncertainty with eq (2.2) it is still necessary that the array be defined. A canasta deck, for example, has two ordinary decks plus four indistinguishable jokers. The uncertainty facing one attempting to guess a card drawn at random is : H = 4/108 (-log₂ 4/108) + 2/108 (-log₂ 2/108) 52 = 0.176 + 5.538 = 5.714 bits/draw. If the jokers of one deck were distinguishable from those of the other, the uncertainty would be : H = log₂ 54 = 5.755 bits/draw. The comparison illustrates two points about uncertainty. First, uncertainty has a maximum value when the items of the array are equally probable (5.755 > 5.714), and second, uncertainty is larger with larger arrays (5.755 > 5.700 of a 52-card deck without jokers). When arrays are of equally probable items, eqs (2.1) and (2.2) deliver the same result, and when arrays contain but a single item (n = 1), uncertainty is zero (no uncertainty if there is no choice).

information

Information is the decrease in uncertainty as the resuit of communication, and hence is calculated as the difference between initial (H₀) and subsequent (H₁)entropy:

with units in bits. Frequently, communication abolishes all uncertainty, so that H₁ = 0, and therefore I = H₀. This is the sense in which authors may use uncertainty (entropy) and information interchangeably

The information associated with communication may be difficult to calculate. Consider surface-feeding ducks of the genus Anas, which have species-specific wing-markings called specula (see figure 20 in Hailman, 1977a). If there are eight equally abundant species on a lake, observing the speculum of a particular individual might be said to communicate I = log₂ 8-0=3 bits/bird. Although this is the information about species transferred to a bird-watcher, it is likely to over-estimate the information transferred to another bird. Each duck probably distinguishes its species from all others combined, so that the initial uncertainty by eq (2.2) is only: H₀ = I = 1/8 (-log₂ 1/8) + 7/8 (-l0g₂ 7/8) = 0.54 bit/bird. In many cases the ethologist has almost no basis for calculating uncertainties at all, and hence cannot estimate the information transferred by a signal.

Haldane and Spurway (1954) cleverly showed that it is sometimes possible to calculate information without being able to calculate a meaningful value for either initial or subsequent uncertainties. The honey bee worker returning to the hive from a foraging site may move in a figure "8" on the vertical comb. The angle of this "dance" with respect to the vertical correlates remarkably well with the direction of the food-source with respect to the sun (von Frisch, 1955). Thus, if the bee moves leftward through the portion of the dance between the two loops of the "8" the source is 90° to the left of the sun; if she moves downward through that segment (the "8" being on its side: ∞), the source is directly away from the sun. Another worker before sensing the dance is initially faced with a full circle of alternative directions in which to search for the food-source, but how many alternative directions are there? If H₀ be measured by the cardinal directions, the uncertainty is log₂ 4=2 bits/foray, but if every degree is recognized as an alternative, the uncertainty is log₂ 360 =8.5 bits/foray.

Haldane and Spurway solved the problem by recognizing that the difference between the logarithms of two numbers may be expressed as the logarithm of the ratio of the numbers: log x - log y = log (x/y). They obtained estimates of the angles at which outgoing worker bees dispersed after sensing the dance, from these data computed the standard deviation SD (which they call the standard error), and then employed an equation of Shannon's to obtain an estimate of Η₁ :

where SD is the standard deviation of a normally distributed population of data and ε is the mathematical constant 2.718. . . , the base of natural logarithms (Shannon and Weaver, 1949). Therefore,

However the standard deviation is expressed (in degrees, grads, radians, etc.), so long as the full circle is expressed in the same units the ratio (circie/SD) is constant and the information transferred has the same value. Gould (1975) points out that the answer obtained by Haldane and Spurway probably underestimates the real amount of information transferred due to technical considerations, but that story is an aside. The example serves to illustrate the usefulness of informational calculations as the "common currency" for comparison among physically dissimilar communication systems. Wilson (1962) later used the same method of calculation to compare dancing of honey bees with chemical trail-following in fire ants.

Usually, the concern is with the average amount of information transferred, but sometimes it is useful to calculate the information transferred by a single type of signal when not all alternative signals transfer the same amount of information. Consider the case in which all uncertainty is abolished by receipt of a signal (I = H₀). The average information transferred (eq 2.2) may be partitioned into its components, each ith signal transferring exactly -log₂ p_i bits, which is weighted in the average according to its probability of occurrence: p_i (-log₂ p_i)· This means that very rare signals (small p_i's) transfer relatively more information at each occurrence than do common ones, since -log₂ p_i is inversely proportional to p_i.

In this sense, information is akin to news as defined in the New York Sun in 1882 by editor Charles A. Dana (or, more likely, his city editor, John B. Bogart): "When a dog bites a man that is not news, but when a man bites a dog that is news." Suppose there are 999 cases a year of dogs biting men and one instance of some canophobe chomping on man's best friend. The average information transferred by reports would have bored Dana and Bogart: I = H₀ – H₁ = 0.999 (-log₂ 0.999) + 0.001 (-log₂ 0.001) - 0 = 0.01 bit/report. The specific report that a man bit a dog, however, transfers 9.67 bits/report. The rumor that the unit of information was named from such animal behavior is, however, without foundation.

It may be confusing logically to use the concept of information in this latter sense applied to individual signals. (Many mathematicians similarly object to the notion of probability applied to individual events.) Attneave (1959: 6) recommends the use of "surprisal" for entropie calculations of individual signals, and in the remainder of this book information implies entropie quantities of signal-sets rather than individual signals.

behavioral outputs and signals

There is no agreed-upon, fundamental unit of behavior, although in many species the behavior appears to be composed of a sequence of natural units called action patterns (see G.W. Barlow, 1977). For a given analytical purpose, behavior may be divided into n distinguishable, mutually exclusive behavioral outputs, denoted b_i (where i = 1, 2, 3, . . . , n). When the number of outputs is specified, the maximum output uncertainty (Η') may be calculated from eq (2.1). Ethologists usually recognize fewer than 50 alternative outputs, and a reasonable upper boundcan probably be set at 1000, so that Η'_max=9.97 bits/ output. If the animal can change output every second, its maximum entropie rate of output could be on the order of 10 bits/s. Information theorists use the same symbols (e.g., Η) to denote uncertainty (entropy) and the entropie rate; the two quantities must be distinguished by context or units (bits/occurrence vs bits/time).

Certain behavioral outputs, particularly certain kinds of movements, postures and orientations of animals, serve as optical signals (see Hailman, 1977a for a review). Moynihan (1970) surveyed the literature on such display outputs and found that the average reported number was about 20 for fishes, birds and mammals; the rhesus macaque had the greatest number (37), which comes as no surprise as it is the most widely studied primate. Wilson (1972) reports that similar surveys of social hymenopterans show that bees and ants have between 10 and 20 communicative outputs. If, say, 50 is a real upper bound, then the maximum entropy of output signals (H*) is on the order of 5.6 bits/output-signal. Of course, there are evident problems with defining the signal-repertoire of any animal, and this question is taken up again in ch 8.

spontaneous transition

From the beginnings of the discipline, ethologists have emphasized the occurrence of spontaneity in behavior (e.g., Lorenz, 1950; Ν. Tinbergen, 1951). Even when the external conditions are held as constant as possible, the animal's behavior changes with time: animals behave without "responding" to anything. For this reason, ethologists rarely refer to the units of behavior as "responses."

The spontaneity of behavior suggests that the performance of an output alters the animal itself in some way, and that internal change of state in turn affects the future outputs of the animal. Control theory provides a framework for treating such a problem. Suppose the animal were a "determinant machine" (Ashby, 1961), such that it always cycled through a fixed sequence of its n outputs. One may tabulate the fixed sequence as a matrix of preceding and next outputs, placing a 1 in cells where the transition occurs and a 0 in cells where it does not (table 2-I). Each column contains exactly one 1 because the control of the machine is determinant. The particular machine illustrated by table 2-I cycles through the outputs 1, 2, 4, 3, 1, 2, . . . indefinitely; its next output is always uniquely determined by its previous output. Such a machine shows Markovian behavior.

Animals are not simple determinant machines; rather, they are probablistic machines whose outputs are rarely uniquely determined. In order to reflect this fact, the entries of the transitional matrix (table 2-I) may be replaced with conditional probabilities. Consider the previous output of the animal to be the measure of its internal state of readiness, or behavioral state, denoted β_j (j = 1, 2, ..., ν). For the moment, j = i and ν = n, although that will change when the notion of β is broadened. The cells of the new matrix (table 2-II) are the conditional probabilities p(b_i|β_j) which may be written more simply p(i|j). As in table 2-I above, the columns of table 2-II sum to unity (the animal must always do something).

Table 2-I

Example of a transition matrix of a simple determinant machine.

Table 2-II

State matrix of a probabilistic machine.

The notion of the behavioral state (β) may now be partially generalized to reflect the diversity of ways in which ethologists operationally judge the animal's state of readiness. One way to achieve higher predictability of the next output might be to consider the behavioral state to be not simply the previous behavioral output, but rather the sequentially ordered pair of preceding outputs. In this case, there are ν = n² columns of the matrix of table 2-II: b₁b₁, b₁b₂, . . . , b₁b_i, . . . , b₁b_n, b₂b₁, b₂b₂, . . . , b_ib₁, . . . , b_nb_n. It is still true that the conditional probabilities in each column must sum to unity, since after each given pair of preceding outputs some next output must always occur. Such a technique may be carried yet another step, by considering the ordered triplet of preceding behavioral outputs to constitute the behavioral state, in which case ν = n³. This technique of probabilistically predicting behavioral outputs from a chain of previous outputs is referred to as sequential analysis or analysis of Markov chains.

semi-Markovian behavior

It is possible to express the output entropy of any column in table 2-II using a modification of eq (2.2) for conditional probabilities:

and to take a weighted average of the column entropies as

By substituting the right side of eq (2.5) for H’_j(i|j),the weighted average can be written parsimoniously as

which may be simplified by noting that p(j) p(i|j)= p(ji):

Equation (2.6) is general for table 2-II, regardless of what is taken to be the behavioral state of the animal.

It is now possible to consider quantitatively the effects of taking longer chains of previous outputs upon predictability of the next output. If the previous ordered pair of outputs is taken to represent the behavioral state, there are nν = n³ cells in the matrix. Let g be the exponent of n so that there are always n^g cells in the matrix. When only the immediately preceding behavioral pattern (output) is taken as the behavioral state, g = 2; when the behavioral state is unknown, or there is only one possible state, g = 1; and when the probabilistic distribution of the n outputs is not specified, let g = 0.

One now indexes the output entropy (H') by the subscript g, so that H'_max calculated before becomes Η'₀; the entropy considering only the overall probabilistic distribution of outputs, regardless of behavioral state, becomes H'₁; the average entropy when the previous output is known becomes Ĥ'₂; the average entropy when the preceding ordered pair of outputs is known becomes Η'₃ ; and so on. An analysis of grooming behavior of house flies by Sustare and Burtt (in prep.) yielded these values based on 14 recognized outputs: H'₀= log₂ 14 = 3.81 bits/output, H'₁ = 3.03, Ĥ'₂ = 1.53 and Ĥ'₃ = 1.33. This result is typical of many ethological studies, in that the largest drop in uncertainty occurs between H'₁ and H'₂ (also see Hailman and Sustare, 1973). Whenever H'₀ ≥ H'₁ ≫ Ĥ'₂≥ Ĥ'₃ , the behavior is semi-Markovian, meaning that the primary observational basis for judging the behavioral state of an animal is its immediately preceding output. The ethological literature suggests that much of animal behavior is semi-Markovian in nature, although communicative behavior may well be an exception.

behavioral states

Two major difficulties with the Markovian approach to behavioral states are the recording and sample-size problems. The approach virtually requires automatic recording methods, which may be difficult to devise and cumbersome to use in the field. If the sequential analysis is to be carried beyond g = 2, tremendous sample sizes are required to fill the cells of the transitional matrix. The second problem is alleviated somewhat when behavior is semi-Markovian, so that there are only n² cells to the matrix, but if the animal has just 10 recognized outputs, there are still 100 cells to be filled with probability estimates,so even in simple cases the problem is not trivial.

Because of these problems, ethologists use other methods for identifying the behavioral state of an animal. One method is to note morphological indicators of internal states, such as engorgement and coloration of sex skin in baboons (see figure 30 in Hailman, 1977a). Another approach to long-term behavioral states in general is the overall complex of outputs being shown by the animal, as when a gull's state is judged as "nesting-building" during the spring period when it is collecting plant material, assembling the material in specific sites, etc. Any indication the ethologist can devise that reflects the internal state of behavioral readiness of an animal usually leads to higher predictability concerning its behavioral outputs. So long as one articulates the list of alternative states, table 2-II may be used to analyze the behavioral data.

When other than transitional criteria are used to judge the behavioral state, table 2-II is no longer a transitional matrix of behavior. It cannot be used, for example, to deduce a sequence of spontaneous behavior under constant conditions of the external environment. For this reason, table 2-II was labeled a State matrix to refleet its general nature. As a non-transitional state matrix, the columns of table 2-II (p. 35) constitute a list of behavioral vectors of probability, denoted β_j. That is, each column is a vector whose members are the probabilities p(b_i|β_j) = p(i|j) and .

inputs and control

The immediate control of behavior in an individual is dictated by internal and external factors; logically there is no third possibility. Let each discriminable combination of external factors be called the external input, denoted E_K (k = 1, 2, 3, . .. , m). Each input specifies a different state matrix like that of table 2-II, and in so doing specifies the behavioral vector of outputs according to the state of the animal at the time of receipt of the input. One way to visualize the conceptual scheme is via a control matrix of internal behavioral states and external sensory inputs, which together specify a particular probabilistic distribution or vector of behavioral probabilities (table 2-III). Or, if the composition of the vectors is to be included, the control matrix has a third dimension of outputs and its cells are the probabilities p(b_z|b_jE_k) = p(b_i|jk) = p(i|jk).

Table 2-III

Control matrix of behavioral vectors.

In order to make the notion of behavioral control precise, it may be expressed as a behavioral control function (C), which specifies a mapping of each pair of behavioral states (βj) and external inputs (E_k) onto a behavioral vector (B_jk) of probabilities of outputs. Symbolically,

where B_jk: p(b_i\jk).

Although the notation of eq (2. 7) is necessarily involved for logical clarity, the idea of behavioral control functions is a straightforward one. For example, Beer (1961, 1962) distinguished outputs of the black-headed gull on the nest:b₁≡ resettling with quivering, and b₂ ≡ resettling without quivering. Two behavioral states were distinguished by the general ongoing behavior of the birds : β₁ ≡ egg-laying state, and β₂ ≡ incubating state. Finally, two external inputs were distinguished: E₁ ≡ one egg in the nest, and E₂ ≡ three eggs (the normal clutch size). Beer created three-egg clutches during laying by adding two eggs to a nest in which the first egg had been laid and created one-egg clutches during incubation by removing two of the eggs from the nest. The probability of b₁--and hence the probability of b₂, which in this case is 1-p₁--was different in all four combinations of input and state. Whether or not a gull would quiver in a particular instance of resettling was never predictable except as a probabilistic statement, which is to say a behavioral vector B_jk.

These external sensory inputs have high diversity. There are so many distinguishable alternative stimulus patterns that can fall upon the retina that one almost believes visual inputs alone to be infinite in number. However, there are certainly limitations on how many inputs the visual system can distinguish as different, and some attempts have been made to determine the entropie rate of input accepted by human subjects viewing random-dot patterns in quick succession. If such experiments were extended to include the color of dots, acoustic inputs, chemical inputs, etc., the average rate of sensory input would be enormous. Consider a vertebrate eye with more than 10⁸ receptor cells. A fanciful calculation that ignores color vision, considers each cell merely to be on or off (no gradations) and considers all spatial patterns created by the retinal stimulation to be distinguishable yields an Input entropy of 'H = log₂ 10⁸ = 26.6 bits/input pattern. Distinguishing a pattern each second provides an input entropie rate of more than 25 bits/s.

behavior

During the foregoing development of behavioral control, it was noted that the maximum output entropy of behavior is on the order of 10 bits/s (probably much less), whereas the input entropy might be on the order of twice that value (probably much more). Similar fanciful calculations may be made concerning the energetic exchange during behavior.

Output energy may be calculated from metabolic rates by assuming a homeotherm burns 5 Calories of carbohydrates per liter of oxygen consumed. A hummingbird at rest thus expends about 10⁴ erg/s, in flight about 10⁵, and a resting cow expends about 10⁷ erg/s. Sensory inputs, however, have little energetic content. For example, a photon of light (ch 3) at 500 nm of wavelength carries about 4 x 10^-24 erg. If three photons are necessary to activate a given photoreceptor sufficiently to cause visual sensation, the energetic input is about 10^-23 erg/cell, and if half the eye's 10⁸ cells are activated during every second, the average requirement is 10^-19 erg/activation. If activation lasts for only 1/1000 s. the eye still requires input energy of only about 10^-16 erg/s. The qualitative point is simply that input sensory energy is far less than output energy expended, even while resting. Sensory energy is not (and could not be) the source of energy expended in behavioral outputs: that source is food.

These comparisons lead to a very general statement about behavior. Behavior is characterized by entropie and energetic transductions by an organism, in which the longterm averages convert high entropic and low energetic sensory inputs into low entropic and high energetic outputs (fig 2-4). I believe that Keith Nelson pointed this out to me years ago, but only recently have I recognized the implication of this relation for entropie exchange during animal communication, discussed in the next major section.

Fig 2-4. Characteristics of behavior include the transduction of high entropic, low energetic inputs into low entropic, high energetic outputs over long-term averages.

outputs as signals as inputs

In generating an output (b) the animal also generates physical disturbances (potential signals). The animal expends considerable energy in generating outputs, but little of this energy goes into the physical disturbances that are potential signals. Some energy is lost in the generating process according to the second law of thermodynamics, other energy is transferred to the environment in ways largely irrelevant to communication, some energy is exchanged in form, and a very small amount of expended energy may be transmitted by the signal. For example, one human output is waving. In raising the arm some energy is lost in the process as heat, some energy is transferred to the environment in the form of kinetic displacement of air molecules, some energy is exchanged for potential energy (and hence regainable as kinetic energy when the arm falls), but little of the animal's energy becomes part of the optical signal of reflected sunlight.

Outputs have diversity, and hence calculable output entropies H'. The maximum output entropy is low, but the maximum entropy of signals (Η*) generated by the outputs is lower(H'_max ≥ H*_max), although this point is not immediately obvious. Outputs are distinguished from one another by differences in the signals they generate, so the set of all signals generated must be a subset of the set of all outputs. If this set-relation were found untrue, certain distinguishable outputs would be logically unrecognized, and the output set would have to be enlarged by further distinctions among outputs.

Every received signal is an external input, but not the reverse, so that signals are a subset of sensory inputs. This means that physical disturbances have no role in communication unless they specify state matrices, which is to say they combine with behavioral states to dictate behavioral vectors of output probabilities according to eq (2.7). But the set of all inputs and the included subset of signal-inputs have diversity, in the sense of having two or more members, each with an associated probability. If *H be the entropy of signal-inputs, it follows from eq (2.1) and the set-relationships that 'H_max ≥ *H_max in bits/ input, and hence also in bits/s.

There comes a point in the construction of broad models of living systems that deduction must give way to reasonable assumptions, based if possible on empirical generalizations. Since such assumptions should be made explicit, I ask acceptance of the reasonableness that longterm entropie averages follow the same relations as maximum entropies; i.e., that

context

It was established in the previous major section (see fig 2-4) that input entropies exceed output entropies in the maximum and on the long-term average. This is a necessary deduction from principles of control theory (Ashby, 1961), since if inputs are to control outputs they must be just as diverse:

Combining this inequality with those of eqs (2.8) and (2.9) leads to the statement

This relation presents no real difficulties for behavioral transduction because input and output entropies can be decomposed into signal and non-signal components. Let nonsignal components be denoted °H and H°; then fig 2-4 may be revised as shown in fig 2-5.

Equation (2.11) does, however, present an apparent paradox for communicative behavior. If the communicants are similar, in the sense that two members of the same species are similar, they have similar entropie relations. However, eq (2.11) shows that the output signal-entropy of the sender (*H) is at best equal to, but usually much less than, the input entropy of the receiver ('H) · In short, signals themselves are not sufficiently diverse to control the communicative behavior of receivers.

Fig 2-5. Partition of entropies of fig 2-4, showing that input entropy ('H) may be decomposed into entropy from signals of other animals (*H) and entropy due to non-signal sources collectively called context (°H). Similarly, output entropy is divided between signals (Η*) and outputs that are not signals (H°).

The point may be made in another way by considering two communicants, A and B. The signal-output of A (H*_A) is at best equal to the signal-input of Β (*Η_B), and usually will be much less because Β will not receive or distinguish all the signals of A. Then B's signal-output entropy (H*_B) will be yet smaller, but by eq (2.11) should exceed A's output-entropy if B's signals are input to A. Then A's output-entropy (H'_A) is larger than A's signal-entropy (H*_A), so that one has come full circle. The paradox is thus that the H*_A of the first exchange of signals is shown to be larger than the H*_A of the next exchange. Were this to be true, entropies would run downhill quickly during life.

The apparent paradox is resolved by recognizing that each communicant must be receiving additional entropy from sources other than signals (i.e., °H), as shown in fig 2-6. The set of all such non-signal inputs may be called the context of communication, the importance of which has been usefully emphasized by W.J. Smith (e.g., 1968).

the "CB" model

This is a favorable juncture for an aside concerning a commonly used ethological model of communication that might be confused with the notions developed in this chapter. This model treats communication as if the animals exchanged signals in some discrete, alternating fashion. For example, A signals to Β, then Β to A, then A to B, and so on. I call this the "citizen-band" (CB) model of communication, in analogy with citizen-band radio in which the sender concludes his broadcast with "over," listens to the other communicant until she says "over," and then sends again.

Fig 2-6. Entropic flow in communication, combining notions of figs 2-1 and 2-5. The Venn diagram at the top indicates the partial overlap of contextual sources of input to two communicants, and what appears to be the same signal (e.g., H*_A and *Η_B) is given two denotations because the signal can be transformed in passage by noise (see fig 2-7, below).

Studies utilizing this CB model construct a transitional matrix of A’s signals and B’s signals, and then perform Markovian analyses using equations similar to those presented in this chapter. The CB model therefore resembles in certain formal and quantitative aspects some of the elements of the conception of communication developed here.

The CB model is, however, quite different. Also, it makes several hidden assumptions about the communicative process. Schleidt (1973) has thoroughly criticized the CB model, and it is necessary only to add a few notes here. As Schleidt points out, the model cannot handle long-term effects of the signals upon the receiver, since the model assays only for the effects upon the immediately returned signal. The model presented in this chapter, on the other hand, conceives of signals as selecting state matrices, which carries no implications whatsoever of an immediate change in output. My model allows for what Schleidt calls "tonic effects" of communication: the maintenance of particular behavioral states in receivers.

Furthermore, the CB model makes no provision for either spontaneity of behavior due to transitions of outputs in the absence of changes in input, or for non-signal inputs of context. Yet another limitation of the CB model is its implicit assumption that signals themselves have no syntactic relations : signals are treated as being read one-by-one by receivers, who reply simply to the last signal received. This problem merges into another one pointed out by Baylis (1976): because the model excludes the possibility that A gives two signals in sequence, the "informational" calculations are misleading. The "information" calculated by the CB model, by the way, does not measure the amount of information being transferred between communicants; at best it measures what the observer learns by watching the communicative interaction. Even this calculation is open to doubt, and Baylis has provided an alternative method for calculating the total amount of information about the system that one obtains by observing it.

Basically, the criticisms amount to this. The CB model assumes that A signals, Β replies, A signals again, and so on. Only in very special circumstances, if ever, is animal communication so simply described, and even when it happens in this way the model for treating it ignores vaiuable data. The purpose here, however, is not to scrutinize the inadequacies of the CB model, but rather to draw attention to its existence so that it will not be confused with the more general conception of communication presented in this chapter.

feedback

When the present value of a variable influences its subsequent value the system containing the variable exhibits feedback. There are three principal sources of feedback by which an animal's output influences subsequent outputs. One of these is represented by the state matrix of table 2-II (p. 35): the present state (β) of an animal determines stochastically its next output (b), which in turn may affect markedly the behavioral state, particularly in semi-Markovian behavior.

The other two routes of feedback in a simple communication system are external. The output-signals of one animal become the input-signals of another, thus affecting its output according to eq (2.7). The output-signals generated by this second animal can in turn act as inputs to the first, thus closing the feedback loop of reciprocal communication.

The final feedback is via context provided by the environment. Non-signal outputs of an animal may affect the environment (some of the output entropy is lost); in effect these outputs act as signal to components of the environment when the components are treated as receivers. In dealing with animal communication (figs 2-2 and 2-3) it is convenient to consider only sender's outputs that are directly inputs to another animal as the "signals." The environment, in turn, provides as its outputs variables that act as inputs to animals, once again closing a feedback loop.

Some ethologists write as if the contextual entropy were shared by the communicants, but this is only partly true. No two animals have precisely the same external inputs, so their contexts will always differ at least slightly. This fact is incorporated into fig 2-6 (p. 45), which diagrams the external feedback loops of entropie flow in a simple communication system.

recognizing communication

The definition of communication as a transfer of information via signals in a channel between sender and receiver still serves well after the development and scrutiny of a broad model of animal communication. The definition does not, however, instruct the observer as to how to recognize communication. The operational instruction of Klopfer and Hatch (1968: 32) has been accepted by others (e.g.״ Schleidt, 1973: 359): "the ultimate criterion for recognition . . . is that of a resultant change, sometimes delayed, sometimes scarcely perceptible, in the probability of subsequent behavior of the other communicant." I should like to offer minor improvements in the instruction. Conceive of two behavioral situations identical in external input (E_k) and behavioral state (β_j) of the reputed receiver, but differing in the fact that the external input of one includes a reputed signal absent in the other case. If the behavioral vector (B_jk)of the reputed receiver differs in the two cases, I snail say that communication has occurred: there has been a transfer of information due to the signal. This is a more precise rendering of the instruction in Hailman (1977a).

It is useful to conclude the discussion of basic communicative behavior with an example that emphasizes the irrelevancy of immediate changes in receiver-output as the criterion for recognizing communication. Suppose a cat is crouched at a mouse-hole. Under constant external conditions, the cat may remain crouched for some time, but as the continual output of crouching alters its internal behavioral state, the cat will eventually "give up" and leave the hole. This change in output is spontaneous, and not well labeled as a "response." Employing the operational instruction for recognizing communication, consider a similar incident in which just before the cat would have left, a mouse in the hole ventures sufficiently far up to be seen by the cat. Instead of leaving, the cat now remains at the hole. However unfortunate for it, the mouse has communicated with the cat, and we as observers have recognized the communication--not as a change in output of the cat (indeed, the effect was to prevent a change in output)--but rather as a redistribution of probabilities of output: a change in behavioral vector.

semantics, syntactics and pragmatics

Peirce's ideas of semiotics were extended by Morris (e.g., 1946) and Cherry (1957) in various ways. Morris appears to have identified the triad of semiotic problems: semantics, syntactics and pragmatics. For purposes of the restricted aims of this volume, semantics may be taken as the study of relations of signals to their referents, Syntactics the study of the relations among signals, and pragmatics the study of the relations of signals to their effects. Weaver, in an insightful essay appended to Shannon's development (Shannon and Weaver, 1949), rephrased the semiotic problems, which he called levels of analysis. Weaver asked how precisely transmitted signals conveyed the sender's intended meaning (semantics), how accurately signals were transmitted (syntactics) and how effectively the received signals affected the receiver's conduct (pragmatics). Cherry's (1957) treatment also differed subtly from Morris's, and appears to have been the primary influence on ethologists who have used the framework (e.g., Marler, 1961; Smith, 1968). Smith is concerned principally with the factors that produce signals, which he calls the "messages" of communication (semantics) and the effects of signals on the behavior of receivers, which he calls the "meanings" of communication (pragmatics). My purpose is not to explore the differences in usages of the semiotic triad, but merely to use a version of it to organize factors relating the kind of information transferred to the design of optical signals (ch 8).

Peirce himself provided the initial classification of semantic relations of signs to their objects, which I use in the physical terms of signals and their referents (pp. 26-27). He divided signs into indexes, which point out their objects; icons, which resemble theirs; and symbols, which stand for theirs in other ways. This traditional classification proves useful in analyzing semantic characteristics of optical signals in ch 8.

deception, distortion and noise

One sometimes thinks of communication as a simple process by which the sender informs the receiver about himself or other phenomena in the world, but the process is plagued with nuances that render it anything but simple. For example, the sender may transmit "misinformation" by promulgating a signal that the receiver interprets as standing for something that it does not. Chapter 6 deals with the problem of deception in terms of interspecific communication as a prelude for considering forms of inter-specific deception in ch 8.

Signals as received are not always identical with signals sent. The simplest kind of change in a physical signal during transmission is distortion, in which the change entails no loss of information transferred. Temperature differences or salinity differences in water may create optical lenses that distort the perceived shapes of fishes and other animals, but the SCUBA diver and probable the animals themselves can often still identify the species-specific characteristics that serve as optical signals.

When the informational content of the signal sent is altered during its transmission, the disturbance is called noise. The notion of noise comes directly from static in radio transmission, where one may mistake parts of the signal sent for something different because of sounds added or changed by electrical interference. Chapter 7 considers kinds of optical noise and their effects on the design of signals.

Fig 2-7. Relation between noise and equivocation, defined in a Venn diagram showing the entropic content of the signal as sent (H*) and as received (*H) in a noisy channel. The intersection is the "useful" information communicated.

Noise is measured by the confusion it causes in the receiver, the quantity Shannon named equivocation. Equivocation is that part of the sender's signal-entropy that is not available to the receiver, and the set-relationships of noise and equivocation may be diagramed as in fig 2-7. In essence, noise is entropy received as signal but generated by some source that the receiver cannot dissociate from the sender.

Overview

Communication is the transfer of information via signals sent in a channel between a sender and a receiver. The occurrence of communication is recognized by a difference in the behavior of the reputed receiver in two situations that differ only in the presence or absence of the reputed signal. The behavioral difference, however, is not necessarily a simple change in what the animal does (behavioral output) just after receiving the signal; rather, the signal causes a redistribution of the probabilities of outputs (a change in behavioral vector). Therefore, the effect of a signal may be to prevent a change in the receiver's output, or to maintain a specific internal behavioral state of readiness. The understanding of communicative behavior requires a synthesis of semiotics, cybernetics and ethology; a preliminary framework is offered. The analysis, of optical signals uses the basic framework to characterize the communicative system: the channel is considered first (ch 3), then the sender (ch 4) and the receiver (ch 5). Deception is the transfer of "misinformation," an important aspect of interspecific communication (ch 6) that has later relevance to social signals as well. Noise is information added to signalinformation in such a way that the receiver cannot dissociate the two, and hence suffers confusion (equivocation). Optical noise plays an important role in visual signaling (ch 7). Finally, the problems of what kind of information is transferred and how it is encoded involves semiotic notions of semantics, syntactics and pragmatics, considered in relation to optical signal-design in ch 8.

Recommended Reading and Reference

Sebeok’s (1968, 1977) volumes offer the best introduction to various views attempting to synthesize semiotics, cybernetics and ethology; Johnston's (1976) recent paper is an important adjunct. The entire September 1972 issue of Scientific American is devoted to communication in its various ramifications.

C.S. Peirce's semiotic ideas may be found in a collection edited by Hartshorne and Weisee (1931-35) and in his letters to Lady Welby (Lieb, 1953). Influential derivative works include Morris (1946) and Cherry (1957), and important contributions from other viewpoints include Chomsky (1957) and Pierce (1961). The foundations of cybernetics are Wiener (1948) and Shannon and Weaver (1949), but Ashby's (1961) treatment of control systems is independent of linear systems of engineering, and Singh's (1966) work on information theory is a fine introduction. Attneave (1959) is another useful introduction to information theory, especially for behavioral scientists.

Tinbergen's (1951) Study of Instinct is a classic of ethology, now seriously outdated, but a thoroughly operational and comprehensive replacement has yet to be written. Hinde (1970) is a useful introduction to problems of control and ontogeny, and Brown (1975) to evolutionary questions about behavior. The interdisciplinary elements that contributed to modern synthetic ethology may be appreciated through Klopfer's (1974) historical treatment.