This paper stems from an experience made by Annalisa D’Andrea (with her colleague Daniela Pavone) as a student of the ’Artificial Intelligence and Intelligent Agents’ graduate course at the University of L’Aquila (held by Prof. Stefania Costantini and Dr. Arianna Tocchio). The experience has been further developed by Annalisa as a Ph.D. student, with her collegue Niva Florio and with the help of Dr. Giovanni De Gasperis.
The application that we illustrate here is developed in DALI, an agent-oriented
logical language that extends Prolog with reactive and proactive features (see, e.g. [
These DALI features allowed us to model a basic element of Pirandello’s plot: along with the interactions, the characters grow more and more suspicious against each other, and their willingness to communicate decreases until finally arriving to a total blackout. This is rendered by modeling trust management in the communication meta-layer, thus subjecting message exchange to the agent’s level of trust in the other agents.
With this example we envisage for future agent-based Artificial Intelligence an extended Turing test, where (in perspective) a human should find it not so easy to understand whether (s)he is talking to a real character or to its agent counterpart. In fact, several students of the Artificial Intelligence and Intelligent Agents graduate course have chosen to model their favorite movie or book in DALI as the subject of their final project, with surprisingly interesting results. In the future we will try to propose “The impossible interviews” (“Le interviste impossibili”), drawing inspiration from a programme dating back to 1975 by RAI2 (one of the three radio networks owned by the Italian government) where a contemporary scholar pretended to interview the “ghost” of some famous person lived in the past. Agents might play the role of the interviewed but possibly also of the interviewer.
The paper is structured as follows. In Section 2 we briefly introduce the DALI language. In Section 3 we summarize some aspects of “Cap and Bells” and how we represent them in DALI. Finally, in Section 4 we conclude. 2
DALI [
External Events. The external events are syntactically indicated by the postfix E. When
an event comes into the agent from its “external world”, the agent can perceive it and
decide to react. Reaction is defined by a reactive rule which has in its head that external
event. The special token :>, used instead of : , indicates that reactive rules performs
forward reasoning. Operationally, incoming events are added to a list called EV and
“consumed” according to the arrival order, unless different priorities are specified.
Priorities are listed in a separate initialization file including various kinds of directives. By
means of these directives the user, in the spirit of [
Actions. Actions are the agent’s way of affecting her environment, possibly in reaction to an external or internal event. In DALI, actions (indicated with postfix A) may have or not preconditions: in the former case, the actions are defined by actions rules, in the latter case they are just action atoms. An action rule is just a plain rule, but in order to emphasize that it is related to an action, we have introduced the new token :<, thus adopting the syntax action :< preconditions. Similarly to events, actions are recorded as past actions.
Internal Events. The internal events define a kind of “individuality” of a DALI agent, making her proactive independently of the environment, of the user and of the other agents, and allowing her to take initiatives, to adopt goals and intentions, to execute plans and to manipulate and revise her knowledge. An internal event is syntactically indicated by the postfix I, and its description is composed of two rules. The first one contains the conditions (knowledge, past events, procedures, etc.) that must be true so that the reaction (in the second rule) may happen. Internal events are automatically attempted with a default frequency customizable by means of directives in the initialization file. The user directives can tune several parameters: at which frequency the agent must attempt the internal events; how many times an agent must react to the internal event (forever, once, twice,. . . ) and when (forever, when some specific triggering conditions occur, . . . ); how long the event must be attempted (until some time, until some terminating conditions, forever). Internal events are treated by the interpreter analogously to external events. A particular kind of internal event is the goal, postfix G, that stops being attempted as soon as it succeeds.
Present Events. When an agent perceives an event from the “external world”, it doesn’t necessarily react to it immediately: (s)he can first reason about the event, before (or instead of) triggering a reaction. At this stage, the event is called present event and is indicated by the suffix N. Past events Past events represent the agent’s “memory”, that makes her capable to perform its future activities while having experience of previous events, and of its own previous conclusions. Each past event has a time-stamp T indicating when the recorded event has happened (i.e., when it has been reacted to). Memory of course is not unlimited, neither conceptually nor practically: it is possible to set, for each event, for how long it has to be kept in memory, or until which expiring condition. In the implementation, past events are kept for a certain default amount of time, that can be modified by the user through a suitable directive in the initialization file. Implicitly, if a second version of the same past event arrives, with a more recent time-stamp, the older event is “overridden”, unless a directive indicates to keep a number of versions. Old versions of past events are placed in a special structure called PNV, as they may have a role in allowing the agent to reason about the past.
The overall structure of a DALI agent, built out of formerly described elements, is depicted in Figure 1, where the communication components are discussed below. 2.2
The DALI communication architecture [
When a message is received, it is examined by the TOLD check layer, which is adaptable to the context and modifiable by the user. This filter checks sender and content of the message, and verifies if the conditions for reception are verified. If the conditions are false, this security level eliminates the supposedly wrong message. A similar check is performed to out-going messages via the TELL layer.
DALI communication filters and meta-reasoning on messages are specified by means of meta-level rules defining the distinguished predicates tell , told and meta (the latter not discussed here). Some standard rules of general use are pre-defined, to which the user can add new others, specific for the application at hand.
Whenever a message is received, coming from a Sender agent, with content part Content , the DALI interpreter automatically looks for a matching told rule. If such a rule is found, the interpreter attempts to prove told (Sender ; Content ). If this goal succeeds, then the message is accepted, and Content is added to the set of the external events incoming into the receiver agent. Otherwise, the message is discarded. Semantically, this can be understood as implicit reflection up to the filter layer, followed by a reflection down to whatever activity the agent was doing, with or without accepting the message. Symmetrically, messages that an agent sends are subjected to a check via tell rules. For every message that is being sent, the interpreter automatically checks whether an applicable tell rule exists. If so, the message is actually sent only if the goal tell (Receiver ; Content ) succeeds. Via tell rules one can manage, e.g., mailing lists. In general, via the TELL/TOLD filters useless message exchange can be considerably reduced.
tell , told and meta rules are contained in a separate file with respect to the main agent program. Thus, they can be changed without affecting or even knowing the DALI code. The FIPA/DALI communication protocol is implemented by means a piece of DALI code including suitable tell =told rules. This code is in turn specified in a separate file that each DALI agent imports as a library, so that the communication protocol can be seen an “input parameter ”of the agent.
The overall DALI communication architecture is aimed at improving elaboration
tolerance where, [
Communication in DALI is in fact devised in order to be elaboration-tolerant with respect to both the protocol, and the filter. 3
As the representation of knowledge and trust plays in general an important role in the interactions among agents, we are going to show in practice how trust can affect the exchange of information among agents. To this purpose, we have defined as a DALI application a small reduction of the Pirandello’s comedy “Il berretto a sonagli” (“Cap and Bells”). Typical for Pirandello is to show how illusion mixes with reality and how people see things in very different way: words are unreliable and everything can be seen at the same time as true and false. In “Il berretto a sonagli” he emphasizes that telling everyone the truth, the sole and cruel truth, regardless of manners or respect and in spite of social habits, will soon produce isolation and a reputation of being mad.
The three main characters of this play have been implemented by three agents of a DALI system: we have not merely reproduced dialogues and lines written by Pirandello, but, based on the dialogues of this play, we have tried to realize agents that mimic the personality of the characters and that reason spontaneously as a spectator would expect.
The title of this play refers to the hat worn by court jesters, that represents the hat of shame to be worn in public. In this comedy, Pirandello pushes the characters into a dilemma with no apparent solution. It is set in Sicily in the early twentieth century and tells the story of Ciampa, a middle-aged bank clerk who lives with his beautiful young wife, Nina, in a small town where sexual infidelity and dishonor are matters of life and death. His boss’s wife, Mrs. Beatrice Fiorica, a jealous and dissatisfied woman, becomes convinced (though without proof) that her husband is sleeping with Nina. She is shocked by the discovery of the betrayal and, thinking only of revenge, issues an official “denunciation” of the alleged affair. Although a police investigation turns up no improprieties (also because the police officer Spano` does not want to stand against Beatrice’s husband, Cavalier Fiorica), Ciampa’s name is still sullied, on the theory that where there is smoke there still must be fire. Ciampa, who paradoxically has always liked to think of himself as the sole voice of reason and sanity in a crazy context, for the sake of respectability does not abandon his wife, but he seriously consider killing her in order to regain his reputation. Luckily for him however, due to Spano`’s deception Beatrice is declared legally insane, and is hospitalized for three months in a mental hospital for acting like a fool.
In the practical demonstration of the application, the first act of the comedy is represented. The involved agents communicate through messages. Each agent adopts a communication strategy implying that messages can be exchanged only with trusted agents (in particular, with those for which the level of trust is more than a minimum). At the beginning, each agent trusts the others. Along with the interactions, the level of trust of each agent towards the others will change, and in our particular case it will decrease. In the end, it will become so low as to prevent any further exchange of messages. Trust management for this example is implemented in DALI in a very simple way by means of the following rules: told (Sender ; Content ) : trust (Sender ; Trust ); Trust > k: tell (Receiver ; Content ) : trust (Receiver ; Trust ); Trust > k:
These rules specify that a certain minimum degree of trust is required in order to enable message exchange. In fact, the first rule accepts reception of incoming messages only if the degree of trust in sender agent is greater than threshold k. The second rule does the same for out-going messages, that are not allowed to be sent to a receiver trusted k or less. Thanks to this mechanism of interaction based on trust, one can leave the agent’s “core” program untouched, and still determine an evolution on the behavior of communicating agents.
For the internal reasoning of each agent we widely used internal events. We do not report the full DALI code here, but we just notice that by means of internal events each agent/character “meditates” about her/his circumstances: Beatrice about how to seek revenge, Ciampa about how to restore his good reputation, Spano` about how to avoid to acquire enemies. Based on their internal conclusions, they will then decide to do something. In particular, their internal reasoning process affects the level of trust in other agents. Also, for instance, a character can at a certain point decide that another one is an enemy. This conclusion directly cuts-off communication (even independently of the present level of trust) whenever the following told rule belongs to the TOLD layer: told (Sender ; Content ) :
not (enemy (Sender )).
The demonstration related to this paper shows examples of communication between agents representing Pirandello’s characters. It should be possible to notice that the interaction is entirely based on the reconstruction in the DALI system of the Pirandello’s character personalities, and that these characteristics determine the generation of events and the corresponding agent answers. 4
The present work can be seen as part of a long-termed experimentation, for which we thank all students of the ’Artificial Intelligence and Intelligent Agents’ course: in fact, via the projects that they have to prepare for their exam, they learn how to program agents and at the same time they help us understand how to improve the DALI framework. The specification of the project (cf. the Course web site http://www.di.univaq.it/stefcost/AI2.htm) requires students to choose a movie or a novel of interest for them, and implement as DALI agents the main characters and their interactions. In the opinion of all involved parties, this is an example of how learning can sometimes be fun!