Science, Values, and Objectivity

Objectifying values in science: A case study

One part of the science of values concerns what individuals or groups value or take an interest in—these are values considered from a first-person point of view. The values can concern anything including morality, aesthetic matters, religion, politics, lifestyle, livelihood, etc. The science of first-person values includes such things as psychological studies of the values of individual people, sociological studies of the values of social groups, and anthropological comparisons of the values of different cultures.

Another part of the science of values concerns what is good for, or promotes the interests of, some individual or group from an external, third-person point of view. The subject whose interests are being studied might or might not internalize the values used in the external evaluation. Examples of the science of third-person values include studies of the value of a college education, of regular visits to the dentist, or of growing and eating organic food. They also include biological studies of what kinds of traits help creatures to survive, reproduce, and generally flourish. This latter example is directly connected to the story I will relate here. This story concerns my own participation in the science of values over the past decade, which grew out of a value-centered theory of biological teleology I developed fifteen years ago.

My story starts in 1991 when I was invited to present that my theory of teleology to the newly founded artificial life research groups at Los Alamos National Laboratory (LANL) and the Santa Fe Institute (SFI). Artificial life is an interdisciplinary endeavor that studies life and life-like processes by simulating them or synthesizing them. Much of this work consists of computer models of processes like the self-organization of simple abstract metabolisms or the evolutionary dynamics of populations of simple self-reproducing automata. Artificial life aims to understand the essential properties of the fundamental processes at work in any possible living system. As Chris Langton once put it, its goal is to understand not “life-as-we-know-it” but “life-as-it-could-be.” Pursuing this goal requires having a general and broad grasp of what life is and could be, so the LANL-SFI artificial life group created a seminar on the nature of life. But the group was unable to formulate an adequate definition. Disappointed, the best they could produce was a list of characteristic hallmarks of life (see Farmer and Belin 1992).

Teleology in one form or another is often considered one of the hallmarks of life (see, e.g., Monod 1971, Mayr 1982), but the notion of teleology is no more self evident than the notion of life itself. So, knowing about my work, the LANL-SFI artificial life group invited me to present to them a philosopher’s perspective on teleology. I knew that the artificial life group did not consider itself to understand a theory fully unless it could see how to implement it in a computer model, so I augmented my presentation with a discussion of how to operationalize the key elements of my theory. My theory of teleology concerns traits that are explained by their value, so operationalizing the theory consisted in figuring out how to determine objectively and impartially when a trait’s value or usefulness explains its continued existence. The “objectifying value” of my title refers to this kind of operationalization.

After my lecture, Norman Packard came up and said that he thought it would be easy to objectify teleology in his computer model of sensory-motor evolution. We worked out the details that night and had our first results the next day (see Bedau and Packard 1992). That was an eye-opening episode that convinced me of the usefulness of operationalizing philosophical theories, whenever possible. This paper describes how I objectified value in biology, illustrates the method in a simple evolutionary system consisting of self-replicating computer programs, and then explains two fruits of this exercise. One concerns Gould and Lewontin’s challenge to adaptationism. The other concerns comparing evolutionary creativity in biological and cultural evolution.

Objectifying teleological explanations in biology

My example of objectifying value in science consists of objectifying the value in a certain kind of biological teleology, specifically, the teleology involved in adaptationist explanations.¹ Traits or behavior that can be explained by reference to the utility of their effects are teleological (telic, for the sake of an end), by my lights.² In ordinary parlance, telic explanations are offered for a wide variety of things. These include such things as the actions of conscious human agents, and the structure and behavior of artifacts designed and used by people. They also include the behavior and structure of biological organisms, as well as certain lower-level components such as genes and also certain higher-level groups such as populations and species. All of these can have telic explanations, and in each case the beneficial effect brought about by the explanandum is an essential part of the explanation. Functionality or adaptiveness is sometimes confused with teleology; the two are related but different. Functional or adaptive behavior is just behavior that is beneficial, that “serves a purpose,” regardless of how it comes about. Telic behavior, on the other hand, is not merely beneficial, does not merely serve a purpose. It occurs specifically because it is beneficial, because it serves a purpose. Telic behavior cannot occur merely accidentally or for some reason unconnected with its utility. Analogous considerations distinguish merely functional or adaptive traits from telic traits.

My concern in the present context is how this framework applies to biological teleology, in particular. A range of behaviors or traits of a given organism at a given time are more or less adaptive. If an organism contains a favorable mutation, the new behavior or trait caused by this mutation might immediately be adaptive or beneficial. But that behavior or trait will not be telic until its utility becomes a causal factor in its continual production. This can happen if its behavior persists through a lineage because of its utility.

What I am describing, of course, is the process by which natural selection produces adaptations. An adaptation is a trait (possibly a kind of behavior) that is produced by the process of natural selection for that trait.³ For example, the whale’s fins are an adaptation for swimming. The trait persists due to natural selection because of its beneficial effects for swimming; this benefit explains why it is a product of natural selection. Although traits of individual organisms are the paradigm example of adaptations, we can apply the notion to higher level entities by averaging over traits and organisms. In particular, below I will talk of genotypes (the complete set of traits in an organism) as adaptations. A genotype is an adaptation if it is persists through the action of natural selection, that is, if on average the individuals with that genotype have been selected for their possession of that genotype, i.e., if the traits in that genotype are adaptations.

The crux of my method for objectifying biological teleology is to observe the extent to which items resist selection pressures, for resistance to selection is evidence of adaptation. Since an item is subjected to selection pressure only when it is active or expressed, I call this evolutionary “activity” information.⁴ Simple bookkeeping collects an historical record of items’ activity—the extent to which items have been subjected to selection pressure, i.e., the extent to which their adaptive value has been tested. The bookkeeping increments an item’s current activity as long as it persists, yielding its cumulative activity. If the item (e.g., gene) is inherited during reproduction, its cumulative activity continues to be incremented by the child’s current activity. In this way our bookkeeping records an item’s cumulative activity over its entire history in the lineage. Cumulative activity sums the extent to which an item has been tested by selection over its evolutionary history.

Every time an item is exposed to natural selection, selection can provide feedback about its adaptive value. Obviously, an item will not continue to be tested by natural selection unless it has passed previous tests. So, the amount that an item has been tested reflects how successfully it has passed the tests. If a sufficiently well tested item persists and spreads through the population, we have positive evidence that it is persisting because of its adaptive value. That is, we have positive evidence that it is an adaptation, that it is telic.

But natural selection is not instantaneous. Repeated trials might be needed to drive out maladaptive items. So exposure to some selection is no proof of being an adaptation. Thus nonadaptive items will generate some “noise” in evolutionary activity data. To gauge resistance to selection we must filter out this nonadaptive noise. We can do so if we first measure how activity will accrue to items persisting due just to nonadaptive factors like random drift or architectural necessity. A general way to measure the expected evolutionary activity of nonadaptive items is to construct a neutral model of the target system: a system that is similar to the target in all relevant respects except that none of the items in it has any adaptive significance. (I give concrete examples below.) The accumulated activity in neutral models provides a no-adaptation null hypothesis for the target system that can be used to screen off nonadaptive noise. If we observe significantly more evolutionary activity in the target system than in its neutral shadow, we know that this “excess” activity cannot be attributed to nonadaptive factors. It must be the result of natural selection, so the items must be adaptations.⁵

An illustration of evolutionary activity of genotypes
I will illustrate the evolutionary activity test for adaptations in Evita, a simple artificial evolving system that “lives” in a computer.⁶ Somewhat analogous to a population of self-replicating strings of biochemical RNA, Evita consists of a population of self-replicating strings of customized assembly language code and residing in a two-dimensional grid of virtual computer memory. The system is initialized with a single self-replicating program. When Evita runs, this ancestral program copies each of its instructions into a neighboring spot on the grid, thereby producing a new copy of the program—its “offspring.” Then this offspring and its parent both start executing, and each makes another copy of itself, creating still more offspring. This process repeats indefinitely. When space in computer memory runs low and offspring cannot find unoccupied neighboring grid locations, the older neighbors are randomly selected and “killed” and the offspring move to the vacated space. Innovations enter the system through point mutations. When a mutation strikes an instruction in a program, the instruction is replaced by another instruction chosen at random. With a moderate mutation rate new kinds of programs are continually spawned.⁷ Many are maladaptive but some reproduce more quickly than their neighbors, and these tend to spread through the population, causing the population of strings to evolve over time.

Evita is explicitly designed so that the programs interact only by competing for space. On average, programs that reproduce faster will supplant their reproducing neighbors. Most significant adaptive events in Evita are changes in reproduction rate, so for present purposes a genotype's fitness can be equated with its reproduction rate. Evita has a clear distinction between genotype and phenotype. A given genotype is simply a string of computer code. If two programs differ in even one instruction they have different genotypes. But two genotypes might produce exactly the same behavior—the same phenotype. If a program includes instructions that never execute, these instructions can mutate freely without affecting the operation of the program. Thus multiple genotypes—without phenotype distinction and so with exactly the same fitness—may then evolve through random genetic drift.

To gather evolutionary activity data in Evita two issues must be settled. First, one must decide which kind of item to observe for adaptations. We will observe whole genotypes. Second, one must operationalize the idea of a genotype’s being tested by natural selection. A plausible measure of this is concentration in the population. The greater the genotype’s concentration, the more feedback that selection provides about how well adapted it is. A genotype’s cumulative evolutionary activity, then, is just the sum of its concentration over time.

In order to discern how much of Evita’s genotype activity can be attributed to the genotypes’ adaptive significance, we create a “neutral shadow” of it (recall the discussion above). The neutral shadow is a population of nominal “programs” with nominal “genotypes” existing at grid locations, reproducing and dieing. These are not genuine programs with genuine genotypes; they contain no actual instructions. Their only properties are their location on the grid, their time of birth, the sequence of reproduction events (if any) they go through, and their time of death.

Each target Evita run has a corresponding neutral shadow.⁸ Certain events in the target cause corresponding events in the shadow, but events in a shadow never affect the target (hence, the ‘shadow’ terminology). The frequency of mutation events in the shadow is copied from the Evita target. Whenever a mutation strikes a shadow “program” it is assigned a new “genotype.” The timing and number of birth and death events in the neutral shadow is also patterned exactly after the target. Shadow children inherit their parent’s “genotype” unless there is a mutation, in which case the shadow child is assigned a new “genotype.” The key difference is that, while natural selection typically affect which target program reproduces, random selection determines which shadow “program” reproduces. So shadow genotypes have no adaptive significance whatsoever; their features like longevity and concentration—and hence their evolutionary activity—cannot be attributed to their adaptive significance. At the same time, by precisely shadowing the births, deaths, and mutations in the target, the neutral shadow shows us the expected evolutionary activity of a genotype in a system exactly like Evita except for being devoid of natural selection. The neutral shadow defines a null hypothesis for the expected evolutionary activity of genotypes affected by only non-adaptive factors such as chance (e.g., random genetic drift) or necessity (e.g., the system's underlying architecture).

Figure 1 about here

Neutral variant genotypes are an exception to this moral, but they prove the rule. Notice that the second fitness jump in Figure 2 corresponds to dense cloud of activity waves. Figure 3 is a blowup of these waves. The genotypes in this cloud differ from each other only by mutations at an unexpressed locus, so they all use exactly the same algorithm. They are neutral variants of one another—different genotypes with exactly the same phenotype. So the neutral variants are one and the same phenotypic adaptation. Each genotypic instances of the phenotype is an adaptation because it is persisting due to its adaptive value.