This paper investigates a speci c type of information that should have an important place in the epistemic state of individual agents, namely their experiences. Just as beliefs, desires, or intentions, experiences should be properly represented, and their speci c role in reasoning and decision processes clearly identi ed. After formally de ning what an experience is, the paper explains in what respect experiences di er from and complement beliefs, and are not just ordinary cases. The added value of experiences in agent reasoning and decision making is then discussed in detail.
The idea of experience is central in di erent human activities such as scienti c
inquiry, learning processes for education, or even in the understanding of art,
as well as the role of shared experience in communication. While it has received
some attention in human sciences [
According to Dewey [
Thus, an experience has two main features that distinguish it from ordinary actions. Indeed, at least some of its consequences: i) are unexpected for the agent, and ii) emotionally signi cant for the agent. Then, the experience is worth being kept in mind by the agent. Strictly speaking, the surprise created by the unexpected consequence of the experience is part of the emotion, which also includes other components that make the consequence relevant for the agent. Thus, while the agent is emotionally a ected by the experience, the experience is associated with some consequences that are unexpected with respect to the agent's beliefs. In the following we concentrate on this latter aspect. As such, an experience constitutes a piece of knowledge which di ers from generic beliefs.
In the following, after having introduced the representation setting used in the paper in Section 2, we formally de ne the notion of experience, and show how it interacts with the notion of belief in Section 3. We then discuss the added values of experiences in reasoning and decision making in Section 4, before indicating lines for further research in the conclusion. 2
We assume a logic (L; CN) where L is a set of all well-formed formulas, CN is a consequence operator, that is a function from 2L to 2L. Indeed, for X L, CN(X) returns all the consequences (under CN) that follow from X. We do not require here any particular property from CN (beyond the assumption that CN(CN(X)) = CN(X)). A notion of consistency is associated with the logic (L; CN):
inconsistent otherwise.
L is consistent in logic (L; CN) i CN(X) 6= L. X is
Regarding the logical language, we assume that it contains two kinds of formulas (L = F [ G with F \ G = ;):
i) F L which represent factual beliefs about what is true in the environment. For instance, `My laptop is heavy' is a factual belief since it concerns a particular instance.
ii) G L contains the generic beliefs about the (default ) laws that hold in the environment. An example of generic beliefs is `The rst generation laptops are heavy'. Unlike the previous example, it concerns a class of laptops and not one particular instance.
Moreover, we shall use a list of atoms V that do not appear in L, and which play the role of names. Finally, we let Ags denote a set of agents.
Action is a key notion in the formulation of experience. Indeed, as we shall see in the next section, an experience amounts to performing an action and observing and evaluating its consequences. An action has a name (e.g., \going to the USA"), it has a set of pre-conditions that should be ful lled before performing the action, and a set of post-conditions which represent the consequences of the action. Finally, an action is always performed by one or more agents. De nition 2 An action is a tuple a = (N; P; C; A) such that: i) N 2 V is the name of the action; ii) P F are the pre-conditions of applicability of the action; iii) C F are the consequences of the action; and iv) A Ags is the set of agents who perform the action.
It is worth mentioning that the same action name may be associated with di erent pre-conditions and/or with di erent post-conditions. Let us consider the action \going to the USA". From one year to another, the visa requirement may change and thus, the pre-conditions are not the same. Similarly, for the same action, the consequences may change. 3
In this section, we focus on one particular agent ag 2 Ags who is equipped with a consistent knowledge base L. An experience of this agent (called actor of the experience) is an action performed by the agent itself. The consequences attached to this action are also assigned and evaluated (positively or negatively) by this agent. Thus, the agent establishes a causality relation between the action and its consequences. Note that in contrast with the use of action in planning, the consequences are not given a priori (as part of the de nition of the action), but identi ed by the actor. In some sense, experiences might feed the libraries of actions that we start with when planning.
An experience is personal to the actor, since the same action made by two di erent agents may not manifest the same consequences for both. Although an agent may report some of their own experiences to another agent, the experience reported does not in general become an experience of the latter agent. Moreover, what makes an action a genuine experience for an agent is the fact that it emotionally a ects the actor positively or negatively, and its consequences are not expected by the actor (the agent either expects exactly the contrary of the perceived consequences, or the agent had no particular expectations a priori).
Thus for instance, our weekly shoppings are certainly not experiences. Assuming that amounts to say that there is nothing peculiar or di erent from a normal visit. However, if one meets the president in the supermarket, then shopping gives birth to an experience since it is unexpected to see the president there.
Besides, experiences can both be made in the past or envisaged in the future. All these remarks lead us to the following de nition.
De nition 3 An experience is a pair e = (y; z) such that: { y is an action (N; P; C; A) such that: a) A = fagg; b) P CN( ) \ F ; c) C F s.t. C 6 CN( ).
{ z = eval(C) s.t. eval : F ! S where S is a scale.
If C = ;, then e is a future experience. Otherwise, it is a past one.
It should be emphasized that the action may be a simple observation (e.g., looking at a sunset), or a physical action (e.g., going to a restaurant). So, you hearing somebody who calls you an idiot counts as an experience. Moreover, the preconditions of the action may be misrepresented or misunderstood by the actor. For example, Peter thought he did not need a business visa for China.
Similarly, the consequences may be wrongly associated to the action. For
instance, one may wrongly think that a train is delayed because of the snow whilst
in reality the train in question has some engine problem. What matters here is
perceived causality [
Note that the condition C 6 CN( ) captures the idea that consequences are unexpected for the actor. The idea is that at least some consequences in C do not follow from the knowledge base of the agent. This covers two cases: { either 9c 2 C such that :c 2 CN( ) { or 9c 2 C, c 62 CN( ) and :c 62 CN( ).
This means that the generic beliefs of the agent were not enough to forecast all the consequences of the action in the circumstances were the action took place. In the last of the two above cases, C is consistent with CN( ), but not deducible from CN( ). In the rst case, the consequences obtained are in contradiction with what was expected. In both cases, the experience brings a supplementary piece of knowledge w.r.t. the generic knowledge, and in case of contradiction with what was expected, the agent learns that some (new) exception do exist with respect to its generic beliefs.
Thus, past experiences are not just ordinary reported cases. They have been singled out for their unexpected consequences that did not leave the agent feeling indi erent when it perceived them. As a consequence, the epistemic state of an agent facing a current situation about which it is trying to draw conclusions, is made of three parts, i) the factual beliefs about the current situation, ii) the generic beliefs applicable to this situation, and iii) the experienced cases that supplement this generic knowledge. Note that the experienced cases if they are su ciently repeated may give rise to new pieces of generic beliefs, or may lead to revise these beliefs. However, experienced cases, as such, are not generic beliefs, and should be handled in a speci c manner.
As just said, experiences supplement generic knowledge, but we have to distinguish between two quite di erent cases: i) the consistency case, and ii) the abnormality case. In both cases, the epistemic state of the agent will be represented by a pair of subsets of possible worlds, say K and E, where K is supposed to account for the worlds that are not impossible according to the agent's generic knowledge (and the factual beliefs about the current situation), or if we prefer the set of interpretations that are consistent with this knowledge, while E gathers the worlds that are guaranteed to be possible in the sense that they have been already encountered by the agent in past experiences.
In the consistency case, we have E K. This corresponds to a situation
of bipolar information [
At this step of the discussion, we will view K and E as binary possibility distributions, i.e., ! 2 K (resp. ! 62 K) means that ! is not impossible (resp. impossible), while ! 2 E means that ! is possible for sure since it was experienced. But ! 62 E does not mean at all that ! is impossible, but just that it has not been experienced yet. As can be seen, the \classical" knowledge representation situation corresponds to the case K = E, where no distinction is made between what is \not impossible" and what is \possible for sure". Making such a distinction is important for acknowledging the speci c nature of experiences with respect to generic knowledge, since agents tend to favor the former when reasoning and making decisions.
The abnormality case is more tricky, since then E \ ( n K) 6= ;, i.e. there is some consequence associated with some experience that is in contradiction with the agent's generic knowledge. As long as the generic knowledge is not revised (in order to recover the consistency condition between E and K), the only way to handle this inconsistency is to consider that in fact K represents the set of all the possible worlds that are normally possible, while there exist some abnormal (exceptional) worlds outside K. This is similar to nonmonotonic reasoning where one has to reason under incomplete factual knowledge with rules having potential exceptions. Here some exceptions factually exist since they were experienced. 4
Di erent forms of reasoning may involve experiences. In particular, one may reason about experiences for making a kind of compilation of a set of experiences. One may also reason about a factual situation on the basis of a set of generic beliefs and of a set of experiences. One may also wonder how to take advantages of both beliefs and experiences in a decision, or in an argumentation process. We brie y examine these three classes of problems. 4.1
There are two situations where it makes sense to aggregate elementary experiences into a compound experience.
The rst one is when the agent performs the same action in very similar circumstances, and perceives the same consequences, then the description of circumstances may be minimally enlarged to cover all the slightly di erent circumstances of the elementary experiences. This includes the case of getting rid of details that are nally found irrelevant for the description of the experience. For instance, an agent gets sick each time after eating a particular dish in di erent restaurants, and nally retains as a \global" experience that eating this dish she gets sick, without no longer associating any particular restaurant to this state of facts.
The second case occurs when the same action that is repeated in the same circumstances has led to di erently perceived consequences. This does not mean that there is any inconsistency: experiences never contradict each other, they just accumulate. Still it is important to keep track of the variability of the consequences by taking as a new consequence the disjunction of the consequences of the elementary experiences. Such situation is encountered in case-based reasoning, where, e.g., two second hand cars that are quite similar in all respects may have been sold to signi cantly di erent prices.
Generally speaking, it is a distinctive feature of positive information to be accumulated disjunctively : the fact that a possible world ! is possible for sure because it is observed does not prevent that another world !0 to be also possible for sure because it is observed later. This contrasts with negative information, which corresponds to the classical view of logical knowledge bases, where the more pieces of the information the agent has, the more restricted the set of worlds remaining possible is (since they are not impossible), as a result of the conjunctive combination of these pieces of information. 4.2
Let us rst observe that from beliefs on may infer new beliefs, and even from
generic beliefs, one may infer new pieces of generic belief, see, e.g., [
Generally speaking, when reasoning about a factual situation s, experiences may either complement or challenge beliefs, as already explained. They complement beliefs when the conclusions obtained from beliefs that apply to s are imprecise or even amount to complete ignorance. However, s may appear in the preconditions of some past experiences. Then, one may take advantage of the consequences of those experiences in order to draw plausible conclusions about s.
Still the problem of reasoning with both beliefs and cases, which requires a
bipolar setting, has received little attention up to very few exceptions [
Example. We have two pieces of generic beliefs (coming from general regulations), namely the two following rules: { \if an employee is in category 1, his monthly salary (in euros) is necessarily in the interval [1000, 2000]", { \if an employee is in category 2, his monthly salary (in euros) is necessarily in the interval [1500, 2500]".
Suppose we also know by experience that { \an employee in category 1 has a monthly salary typically in the interval [1500,1800]", { \an employee in category 2 has a monthly salary typically in the interval [1700, 2000]".
Strictly speaking, all the values in [1500,1800] or in [1700, 2000] have not been observed for employees of categories 1 and 2 respectively, but one may consider that these intervals are the convex hulls of set of values that have been experienced for such employees. Then suppose we have the imprecise piece of factual information that \Peter is in category 1 or 2". What can be concluded about Peter's salary? Using the two generic rules we get that his salary is in [1000; 2000] [ [1500; 2500] = [1000; 2500]. Indeed, from \if p1 then q1", \if p2 then q2", \p1 or p2", one concludes \q1 or q2" (or if prefer in propositional logic terms :p1 _ q1; :p2 _ q2; p1 _ p2 ` q1 _ q2).
Besides, the values in [1500; 1800] \ [1700; 2000] = [1700; 1800] are for sure
possible for Peter's salary. The use of the intersection \ here may be justi ed
intuitively by the fact that the more imprecise the situation, the smaller the
associated set of values that are possible for sure. Indeed, it is only those values
that are common to the two contexts `category 1' and `category 2' that can be
retained as values that are possible for sure in the imprecise context `category
1 or category 2'. A formal logical mechanism that can handle inferences with
positive information can be found in [
When an experience leads to consequences that are not consistent with the conclusions obtained from beliefs about s, we have to give priority to experiences over beliefs (at least if there is no uncertainty on the consequences associated with experiences). This also means that we need a case-based reasoning mechanisms working together with a non-monotonic mechanism for handling the default rules expressing generic beliefs. Something that has still to be done.
Besides, the fact that agents possess both beliefs and experiences should also considerably enrich argumentation, since experiences may provide a natural basis for counter-argumention, by o ering counter-examples to generic rules that have been used in a previous argument for supporting a conclusion, which then becomes challenged. 4.3
Decision under uncertainty takes advantage of the available knowledge (possibly
pervaded with uncertainty) about what can be deduced about the state of the
world in the current situation, and of how good or bad are the consequences
of actions performed in precisely known states of the world. Then, depending
on what are the rationality postulates, di erent decision criteria, such as, for
example, the expected utility, can be justi ed ; see, e.g. [
In contrast with such approaches, Gilboa and Schmeidler have already
advocated the need for making decisions on the basis of experiences, called cases
in [
A formal parallel between decision under uncertainty and case-based
decision has been already made for pessimistic and optimistic qualitative decision
criteria [
Such decision criteria could be adapted when we have both generic beliefs and experiences about candidate decisions. Consider the pessimistic point of view, in the case where experiences are consistent with beliefs (i.e., E K in the notations of Section 3). Then asking for having all cases similar to the current situation leading to a good consequence is less requiring than asking that all non impossible states compatible with the ill-known situation lead to a good consequence. Note that with the optimistic point of view, this is the converse: it is more requiring to have at least one case similar to the current situation leading to a good consequence rather than having one state compatible with the ill-known situation that leads to a good consequence (since the set of worlds similar to E remains strictly included in K).
When some experiences contradict generic beliefs (i.e. in the abnormality case of Section 3), then the agent has to assume that the generic knowledge covers normal situations only, and that some abnormal cases may be encountered as well by experience. Then a more di cult situation is when all normal situations lead to good consequences for a candidate decision, while it has been observed that this decision in some abnormal cases led to bad consequences. If the agent is moderately pessimistic, it may be already satis ed if in all normal states compatible with the ill-known situation, the decision leads to a good consequence, even if some (but not all) experiences where the decision was experimented led to bad consequences. But, if the agent is very pessimistic, the agent will be afraid to make a choice that once led to a bad consequence, even it was abnormal. This suggests the need for designing sophisticated decision criteria able to cover all these situations. 5
This paper has presented a rst manifesto in favor of a forgotten component in
the modeling of the epistemic states of an agent in Arti cial Intelligence: the
experiences held by the agent. Indeed, we believe that experiences are a
component which is fundamentally di erent from beliefs, desires and intentions, and
which deserves to be formalized and clearly distinguished from the others. We
have explained in detail why beliefs and experiences are not of the same nature,
and we have indicated in general terms how they may impact both reasoning
and decision making in general. This should lead to a new generation of agent
architecture, that we may call XBDI agents (where X stands for experience),
which enlarges the original BDI architecture [
The new architecture will make it possible to de ne platforms that allow several agents to share experiences. For instance, one may imagine an application in which agents bring their past experiences to others about a particular topic. We can also think of another type of application in which a group of agents want to have a joint future experience such as visiting a place of interest together, for instance. In this case, the future experience is a common desire for the agents. Both these examples are only possible if the notion of experience for agents is clearly de ned. 6
The authors acknowledge the support of the ACE project (Autonomic Software Engineering for online cultural experiences) of the European Chist-Era programme.