In search of a principled theory of the ‘value’ of knowledge

What is mind for?

Doxastic/epistemic information (Knowledge) is not the center, the end, the sense, and the real nature of ‘Mind’. The center of gravity of mind are Goals, which—on the basis of knowledge—have to successfully drive our behavior. Mind is a system for teleological ‘controlling’ conduct on the basis of ‘representations’ and their manipulation (assumptions on the current state of the world, on the powers of the Agent, future-augmented reality-imagination & anticipation-, desired states,…); a system built to ‘solve problems’ by working on the representation of them (that is, ‘mentally’), by reasoning, planning and deciding.

Our thesis

This distortion in the view of Cognition is the reason why a theory of the Value/Utility of information in term of knowledge is not really there; systematically grounded and developed. This would require (as we said) a view of the role that K has in relation to Goals and their dynamics, management. K value is a notion relative to Goals and their Value. Since Inf/K is a means, the value of means depends on the value of their possible goals/uses. We search for, acquire, buy, preserve, use, consume,.. exchange… this crucial ‘power’² for realizing goals.

The thesis of this paper (a preliminary exploration, not a complete systematization) is that Ks are just a fundamental resource; just means for (potential) goals. Ks have a Value and Utility, they can be more or less ‘precious’; they have a cost and imply some risks; they can not only be useful but negative and dangerous,… In a sense, one should apply to K—in this goal-oriented perspective—a ‘economic’ frame.

Of course, it is true that in humans the acquisition and storage of K has become an end in itself (and even an ‘intrinsic motivation’ and a ‘value’) (section “Goal processing and the utility of beliefs”). Subjectively we do not necessary search for knowledge acquisition instrumentally to a foreseen ‘use’ of it. Nevertheless, the function of K is instrumental. In human evolution, psychology, and society means becomes ends (this holds not only for K, but for social image and relations, for money, for power, for conformity to norms, etc.).

‘Relevance’ is not so clear and is not enough

It is not our aim here to discuss or develop the theory of ‘Relevance’; which, on one side, it is strongly bounded to pragmatics and communication theory (which are not our topics here); on the other side, it is quite rich and complex and not so clearly defined and systematized. We will give here a reductive view of ‘Relevance’ just in order to introduce some important ‘distinguo’. We do not need just the notion of ‘Relevance’ but those of ‘Utility’ and its measure, of ‘pertinence, etc. And we do not just focus on communication or mere cognitive processing, but on cognition for motivated action.

We have two main criticisms towards the Relevance notion and theory.

First, this notion doesn’t have a convergent definition and use,³ and even in the different domains and approaches it is not very clear and rigorously defined; it is based on rather vague notions.

Not just ‘information’

Let’s shortly present the important contribution of Luciano Floridi on informativeness and Relevance, as grounded on probability and information theory.⁷

Floridi clearly enounces the problem, as: “When is information relevant? How can we analyze the concept of information that interests ⁸ the agent?”

This is true and relevant, however—in our view—it would be also required an explicit theory of why information is needed “in order to interact successfully” with the environment; that is, to act and to achieve goals. In his much more advanced model, Floridi however does not abandon the strictly “cognitive” view of Inf relevance and of its “epistemic utility”. The basic notions of the process model still are “question” “answer” etc. (see Fig. 1). It is not really explained what does it mean that a given input “matters” for the agent. His view of cognition seems putting aside an explicit/systematic theory of the motivational aspects and of the action regulation.

Not just ‘information’ quantity

In general, in our perspective, “Information theory” cannot solve the problem of the Utility of the information items.

The mathematical theory of information is based on probability theory and statistics, and it also provides a measure of information, with several quantities of information (Shannon, Kolmogorov,..)

For us this is quite reductive and wrong.

Epistemic utility theory

So the aim of this strong approach is the formulation of norms for a rational ¹¹ believing and truth approximation; and notice that the ‘utility’ of an epistemic representation is just identified with the “truth” or better with: the aim of approximating the truth.

Not just ‘economics’ of information

The ‘Economics’ of information typically focus on rather different problems.

Network-based Pertinence

Reasonably, there is also another crucial factor for determining and predicting if a given K/D/belief is informing, is pertinent about X. It is the structure of the network of (episodic or generic) Ks ‘about’. Our representations—in particular the doxastic ones (data, beliefs,..)—are not stored as files or lists; they are organized in a specific way. Let us simplify by using the more clear and schematic ‘propositional’ format. The proposition (content of a possible belief) P’ is represented in terms of a given predicate (Pred) with its argument(s) (arg). For example B1: Pred (arg1 arg2); B1: Eat (John, bred). The organization of the data/Ks is around its arguments. We put around arg1 (John) all the beliefs or data we have directly about it, that is, mentioning it as an argument; and the same for all the data about arg2. If the predicate is not a simple single-argument ‘property’ but is a ‘relation (like ‘to eat’) this proposition (belief,..) connects the two (or more) nodes, and these representations and Ks, builds a Network.

Such network determines relation of direct vs. indirect connections, and of proximity; and also degrees of connection. For example, given the network in Fig. 2a, X is clearly closer to, more strongly and directly connected with Y than with Z. However, not only the ‘steps’ matters, but also the number of connections, the topology of the net. For example in Fig, 2b X is more connected with Y then with Z.

However, in our view this relation is not so trivial: Is necessarily informative about X any K about/on Y if Y is strongly connected with X? Not necessarily; not all the Ks about Y are pertinent for X. John has a dog, this dog yesterday has meet another dog in a garden; is that pertinent and informative about John? Non necessarily. But of course—in our view—even less informative about X is a belief about an entity not connected at all with X; a dog in another continent and ignored by John.

We remain on our claim: to be pertinent for X, to give information about X, a K not directly about X must potentially be the basis, must give the opportunity of inferring something on X. The relation with the network structure and its dimensions (proximity, number of connections) is this one: the strongly related, connected Y to X, the more probable that a D/K on Y be informative (has something to say, to derive) about X. If Johns’ dog gets fleas this probably has consequences for John, and we can infer, expect something about him; if a dog met in the garden has fleas this is for sure less pertinent and relevant about John.

So, the structure of connection between nodes is important, but the pertinence and possible informativeness about a node depends on possible inferences.

Another network structure is obviously crucial. In fact it is not just a matter of topology (proximity, etc.), it is also a matter of the semantics of the specific link. A very crucial link is the ISA relation: the link between a Class and its sub-classes and members (individual instances). Around the Class nodes we have the ‘generalized’ K (what we know about not a specific dog or car or action of eating but about dogs, cars, eating,..); this generalized information is there precisely in order to be inferred/instantiated on the members. So, if we (come to) know that X belongs to the Class C’, we automatically (come to) know a lot of things ‘about’ X.

Similarly, our K can be organized and assembled in ‘domains’ (like “sport”, “politics”, “art”,…). If my Obj of interest is within a given ‘domain’, probably the Inf within the same domain can be ‘indirectly’ pertinent for Obj. Or better, the probability that an Inf item in that domain be pertinent (and relevant) are greater than the probability of an Inf item of an independent/unconnected ‘domain’.

Goals¹⁵

A goal is not a special kind of representation, different from a doxastic representation; it is instead a mental representation with a specific ‘use’, ‘function’, ‘role’, ‘application’. Beliefs (etc.) and goals are just one and the same kind of representations, simply employed in two different ways. In fact, it has to be possible for such representations to match, in order to characterize the state in which an agent believes his/her goal to be satisfied. Beliefs and goals are two different possible ‘attitudes’ on the representations. For example, the same visual image can be used in a given circumstance as a belief about the current state of the world, and in another case as representing a state of the world to be achieved, i.e. a goal.

A goal is not necessarily pursued; a representation does not acquire the status of a goal only if and when it is being pursued as an ‘objective’.¹⁶ It is a goal (i.e., it plays the role of a goal) also in other stages of the control cycle; this is in fact the model of modern cognitive notion of ‘goal’ and of goal-directed behavior. Only in some cases, I have to ‘actively’ pursue my goals. When the realization is (believed to be) possible and depending on me, being up to me, then, either I subjectively “try” (the result is not subjectively sure) or I intentionally act for realizing it and confidently expect the desired outcome.¹⁷

In sum, a goal is a goal even before or without being pursued: satisfaction or happiness are due to goal-realization, and dissatisfaction or suffering are due to goal frustration, but not necessarily to our active successes or failures; we are sad because our mother has died, or happy because she gave us a kiss, even if we did not ask for it, nor do or expected anything from her.

There are different kinds of mental goals: end goals vs. instrumental goals (means) and their hierarchical structure; plans; intentions; motivating and non-motivating goals; projects; expectations,.. Let’s just focus on a difference that is relevant for the theory of goal Value: the difference between ‘felt’ and ‘non-felt’ goals.

The origin of goal value

Since we have more than one goal and several goals can be active in the same time and context, and be in conflict (incompatible), this is why we have to ‘choose’, to have ‘preferences’. But this implies that goals have to have a subjective ‘value’, that is, a value perceived or assigned by the decision-maker: a quantitative dimension enabling the decision maker to give them some ordering, by making them preferable or not to other goals.²² Our issue here is the following one: Where does the value of a goal come from?

Deriving the K value

This is a bit too simple; the theory of Value presupposes the theory of ‘evaluation’. Given that we do not have only one goal but open sets of goals; and we may have a set of currently ‘active’ goals, and a set of possible or future goals, we have to consider the Utility of a given Inf or K in relation not just a single (and actually pursued or taken into account) goal, but in relation to a given Set of goals (Sg) considered by the subject in its evaluation. Thus, the value of K is due to the value/importance of the goal-set Sg it is useful for, and their value.

The evaluation (attribution of ‘Value’) to a given entity Obj in fact is usually multi-dimensional; there is not just one parameter, one standard ²⁶ or criteria (in other terms one current or possible use and goal) relative to whom we appraise the correspondence and utility of Obj. This implies also that there is a sum of partial values of Obj. It can be ‘good’ in relation to goal or Value G1 & G2 & G3, and thus be more precious of an Obj just useful/good for G2. However, it can also be very good for G1 but frustrating, harming G2. Thus the Value of Obj is the value of G1 minus the Value of the frustrated goal: it is partially good and partially bad.

Other distinctions

K quality

Our doxastic representations have a ‘quality’: a subjective ‘certainty’. We are more or less “sure”, “convinced” that P.

The degree of certainty should affect the KV: for example, I should ‘pay’ a given K from a very trustworthy source (or a K with a given degree of certainty) more than from a not so reliable source or a doubtful K. The degree of certainty (quality of K) has a important value since we actually bet and risk on that; we decide to spend our resources and actions on the ground of what we believe; so our trust in what we believe exposes ourselves to failure, harms, …

K is a resource that changes its value on the basis of its origin, of the brand, of ‘seller or producer’; or because the source is/gives a value in itself (like in dogmatic knowledge) or because the quality of its products has been proved superior (previous experience, reputation, marketing,..).

In sum, the Value of K is affected by its degree of certainty: the more sure, grounded, the more precious it is.

The value of uncertain and doubtful information is clearly inferior.

The epistemic integration value of K items

There is also another utility and value of Inf items (data, candidate ‘beliefs’), not directly relative to specific ‘motives’, neither as frustrating/satisfying, nor as tool for realizing the goal. There is an ‘importance’ or ‘value’ of a given data or beliefs just in relation to knowledge organization, integration, mutual consistency and support. In a sense this importance or value is due to the ‘pseudo-goal’ of having robust, integrated, knowledge. To have coherent and justified knowledge is a crucial ‘function’ of our cognitive system (for example Thagard 2000).

Any process of knowledge acquisition, generation or elaboration does not only generate some output knowledge item; it also generates a trace of its origin, and ‘relation’ that supports and integrates such a new item. It generates at the same time knowledge structure (network) or ‘relational knowledge’: Reasons to believe.

This is true not only for inferences, but in general. Knowledge items remain related to their source: “I saw that p”; “I think that p, because…” “The TV said that p”, etc. Thus, there is a special relation between the Belief that p, and the Belief “I saw that p” or “the TV said that p” or “Since Q then P”.

We talk about ‘support’ relations, because cognitive items hold thanks to such relations.

Now, this cognitive ‘need’ of reason to believe, of support and integration of K, this structure gives to K items different ‘role’ and ‘importance’ or ‘value’. Not all items are equally important in a given domain, context, or episode (independently on their degree of certainty and on goals). There are K items more ‘central’, ‘important’, ‘crucial’ while other items are just ‘marginal’, just ‘details’. This depends on their network role: is this piece of K supporting and explaining many other Ks of that episode or domain? What earthquake would happen if this K would result wrong? How much belief-revision work we should do? Or this information is quite irrelevant, it doesn’t support or explain nothing, and we can cheaply abandon (drop, revise, forget) it?

In other words, K items have different value and utility in relation to the need for coherence, support, and argumentation within our beliefs. It is more probable (ad reasonable) that we forget or put aside marginal details (that is, facts that are not important for understanding the whole, that do not explain the other facts and the global event) then central facts. It is more probable that we resist more to revise and abandon crucial, central, important items than irrelevant, marginal one.

In sum, K has a peculiar form of Utility and Value—its importance—just relative to doxastic (not motivational) aspects: how much it is integrated in the K nets and how many K items does it support or is supported by, and how central is in the topology of the Net?

Goal processing and the utility of beliefs³⁴

But this also gives Value to beliefs, make them more or less important from another point of view (different from the previous ones).

Beliefs support goals, by acting as reasons for them. A cognitive agent is an agent who grounds his actions on his beliefs: or better, he acts on the basis of what he wants and prefers (goals), but he wants and prefers on the basis of what he believes. In a cognitive agent, goals should be supported and justified by reasons (not necessarily unbiased and “rational”). We activate, maintain, decide about, prefer, plan for, and pursue, goals that are grounded on pertinent beliefs (supported by other beliefs, etc.). Goal-processing is belief-based.

Let’s summarize the role of different types of beliefs in filtering the goals and in regulating their transition step by step, from their activation to the action execution (for extended discussion, see Castelfranchi and Paglieri 2007).

Goal processing has in fact different phases, like: goal activation; choice between various active and competing goals; goal planning and formulation of an Intention to do something; execution of the planned action. Now, all these phases require specific information input and exploit specific beliefs. So, a given belief can be the condition for the activation of a given desire (for example, I see that there is an open ice-cream shop in front of me); or a belief is the condition for abandoning a given goal (the believe that it is impossible or the belief that between two active goals there is a contradiction, a conflict, and we have to choose between them); or a condition for choosing G’ and not G2 (for example, I believe that this dress that I personally like less that the other will be more appreciated by John and I want to like to John); or a condition for formulating the intention and the planned action (for example, to believe that such action has those results and that I’m able to perform it); or a condition for the action execution (for example, the information that it is the right time). Failing these belief tests would stop the processing of a goal (e.g., putting it in a sort of mental waiting room, until the agent beliefs will allow reactivating them), or may even eliminate certain goals; on the contrary, success of these tests will make a certain goal persist until the choice, the planning, the execution of the pertinent actions.³⁵

Now, given that a given belief B’ is responsible for the acceptance of that G as ‘to be possibly pursued’ (up to me), or of the choice of that G as better than its competitors, or as a possible ‘intention’ of mine (I’m able and in condition) etc. it acquires a special ‘supporting’ Value. If B’ is revised, I have to drop that G; and this may be a cost, a waste; or I don’t want to renounce to G for its importance and attractiveness. This crucial role of B’ may entail unconscious ‘refusal’ to revise it³⁶ and even self-deception in order do not renounce to my G or do not frustrated it.

To know just for to know

We have (for evolutionary and cultural reasons) the final Goal of acquiring knowledge, the intrinsic motivation of curiosity. Thus knowledge and its acquisition can be/have a Value per sé.

However, first, curiosity is not for everything, omni-directional (at least, it is focused on certain domains: our ‘interests’, in order to restrict an infinite search). But this means that we are ‘interested’ in that, that is, we have the goal to acquire K in that domain, on that topic.

Second; it is true: we have the final goal to ‘cumulate’ and ‘store’ knowledge also without a clear immediate use of; like for money, or like ants for food. K it is a ‘good’ per sé. However, not all the founded Inf has the same value. Like for money it is different to put away 1000§ or 10§. How do we evaluate the value of a storable K item, which is not immediately to be used for a given active goal G’?

The generic goal to acquire knowledge is insufficient for explaining that. We plausibly evaluate in an approximate and intuitive way the presumable Value of that new information either on the basis of our ‘interest’ in the domain/topic it is Pertinent for (an area we know that we will have goals of specific uses), or in term of potential possible uses/goals (section “Other distinctions”). Like when I decide to buy a dress that I do not need just now for a specific ceremony (or I do not want it will be used only for that) and I evaluate the probability of other future occasions and good uses of it: “this dress is better; I’m sure that I can use it for a lot of other occasions and roles”.

So, even for the INF/K acquired just for the goal of acquiring K, there are additional Values/Utilities relative to not just to ‘know’, but to possible uses, interests.

Value losing

The costs and risks of K

A piece of K also has costs.

(i) Acquisition costs:³⁷

How much one has to invest (or have invested) in order to have K; in terms of time, effort, material resources (money,..), relations, sufferance,….

Obviously, the cost should be acceptable relative to KV current or possible or for multiple uses.

Nice examples of costs for acquiring K, are: to study; to pay an investigator; an intelligence service; reasoning and calculating; to memorize and store and to try to retrieve, etc.

It is also very relevant the relationship between the costs of K acquisition and the risks.

‘Knowledge’ in strict sense and the dangers of truth

We are sorry for Mill (1823) but there can absolutely be ‘noxious’ K; not all K is by definition good for humans, as he claims. This is even clearer if the parameter is not our “good” but our subjective “happiness” (not the same!).

Relative to our “happiness” how many times we sincerely say “I would have absolutely preferred do not know that! I suffer a lot and cannot do anything at all!”. Or even, what I now know obliges me to decisions that waste my life and I wouldn’t like to take. I would prefer to live in ignorance or deception.

We would like to avoid a deep discussion about what is “knowledge” in strict sense (true epistemic representations). As we said we use K as a broad notion (like used in AI, cognitive science) covering different doxastic and epistemic representation: Information, data candidate to be assumed or not as belief or K, beliefs with different degree of certainty, presuppositions, as-if assumptions, hypothesis, expectations,….).

Our view of “truth” is radically pragmatic (related to the theory of action and goals): that a given assumption, presupposition, explicit belief, subjective ‘knowledge’³⁹ is efficacy; that is, that an action grounded on such a belief (on the world, on me, etc.) and relying on such represented ‘state on world’ be successful (goal achievement). A “true” knowledge is—in this radically pragmatistic sense—intrinsically ‘useful’ and of Value, since it is the basis of the action and of its success. “To be true” is a value and standard, and gives value. Nevertheless, consider that there is not only one goal in our mind and in our ‘evaluations’. For example, to “use/rely on” that assumption and to “verify” it, are not one and the same goal. The belief in the truth of a given K item, implies the possible action and potential goal of “verification”, an epistemic action aimed at looking at, and seeing if, and verifying or falsifying if P. Also relatively to that goal (motivating that action: that our assumptions be verified) that K is more or less ‘useful’, not only in relation to the ‘practical’ goal of achieving G’. And the results, success, could be different.⁴⁰

Also for this radically pragmatistic notion of K, it follows that any K (as such) has value, since we have the goal of the truth and that what we believe be true⁴¹; but it doesn’t follow that such a INF item since it is true (K) cannot be useless, or even harmful and with a negative value relatively to other goals of us. Value is always in relation to several goals and dimensions; and that Inf Item (although true) can be ambivalent: good as for the goal of truth but bad as for other goals or functions.

It seems to me that this view is not contradictory with a pragmatic definition of truth; the item could remain true even conducing us to some failure on some goal. It is successful, valid, for certain actions but not for others or for additional goals it elicits or for bad inferences it implies.

Let’s say something not in relation to our (rather preliminary) view of “knowledge” in strict sense, and of truthfulness, but in relation to the classical view of truth: the assumption correspond to the ‘real’, objective (subject independent) state on the world. Like in logical definition of the operator “to Know”, where (Know X P) = (Believes X P) & P. “X knows P” means that “X believes that P” and “P is true” (from the perspective of an ‘observer’ assumed as the real state of the world).

As for (A) the answer is: It depends! “Good” in relation to which set of goals, to which evaluative dimensions? As we said, for the goal of known the truth, that K Item has always value, it contributes to; but relative to other goals it is pertinent and relevant for, it can be useless and even negative, harmful; although ‘true’.

Thus, a partially unrealistic and incorrect Inf and beliefs can help.

More in general, also at the social-political level there is a utility of illusory⁴³ and utopian beliefs. They play a parabolic role for human ambitions and objectives: in order to rich X we have to aim to in Y.

“By striving to do the impossible, man has always achieved what is possible. Those who have cautiously done no more than they believed possible have never taken a single step forward.” Michail Bakunin

Deciding to believe?

First of all, the basic mechanism/process of ‘deciding’ to believe or better to ‘arriving’ to believe is not a true ‘decision’, based on advantages and convenience; I like to believe that P so I decide to believe it. This is in case just a deceptive and unconscious process of ‘motivated reasoning’, or a ‘defensive mechanism’ (Miceli and Castelfranchi 1998). We cannot for example believe that something is true just because X pays us if we believe so, or because X would harm us if we do not believe so. ‘Economic’, rational, convenience doesn’t govern the believing ‘decision’; we can just declare to X and simulate that we believe what he likes. Believing process has its own bases: source credibility, supports, evidences (section “The value of knowledge and its origin”); it is an epistemic process not a utility driven one. And the convenience of a given belief cannot (consciously) alter its ‘credibility’.

Promises, threats, rewards, are useless for inducing to believe. This is one of the guaranties of our ‘autonomy’ and defense from social manipulation.⁴⁵ Thus, we do not believe something (D/K) because it is convenient for us to believe it. However, to believe it is convenient and makes D/K convenient, useful (or dangerous). Moreover, as we said, the degree of certainty, how much we believe in D/K, increases the subjective value we ascribe to it. We feel safer while betting on it, and we preserve it strongly.

Let’s assume that P is good, useful for X; then also knowing (believing) P is useful for X. If a given K has value for me and is objectively useful for me, it is convenient to me know it and believe it; but I do not believe it because I think that it is convenient to me believe it: I believe that it is probably true, reasonably grounded, and useful for my goals.

The value of information for our image asset

A really special value acquire Inf Items about us and our behavior and this is why we want the other know them; or—vice versa—the relevance of the ‘secretes’ about us, what we hidden and don’t want the other know. The function of this maneuver and the positive/negative value of that Inf Items is the shaping of our social ‘image’ and of evaluation about us, then of our esteem, trust, reputation,..

Our comparative evaluation (due to the good information about us and the hidden of the bad ones) is also vital for us because it is the substance of human social (non formal) hierarchies, and of our positioning in them (prestige, visibility, and relational capital, opportunities..). We socially live of what the other know and believe about us.

To share or not to share K? Effects on its value

Let’s give two examples of possible reasons for K sharing:

B) Intimacy: K to be shared and not to be shared A very strange/remarkable form of shared/not shared K, with its ‘rules’ in order to be and work ‘as’ “intimacy”, are those Ks about our body, our thinking and feelings, our story and narrative, that not everybody has to have access to (to see, to listen to,..). Otherwise I would not have ‘intimacy relations’ and ‘intimacy’ parts of me.

So, the rule is that there are people that has not to know (see,..) those Inf about “me”; I have to ‘protect’ and hidden them (decency,..); while there are other people (my intimate relations) that has to known; I have to share these K with them, otherwise they are not really in ‘intimacy’ relation with me.

And this adds, gives special (social) Value to those Ks.