Discreet Intelligent Agents

With the development of information technologies, the world becomes so complex that it is now necessary to have many virtual robots – i.e. webbots, search engines, etc.   – at our disposal to alleviate our cognitive overload.   These robots are not androids, like many nonhuman inhabitants of   science fiction movies, but virtual, i.e. pieces of software who behave like humans, or more generally like agents. The virtual robots – or intelligent agents – have to act as intermediaries between us an the technologies. By reference to the etymological meaning of the word “robot”, virtual robots can be seen as artificial workers that will certainly play a central role in the near future everyday life. For instance, they will facilitate shopping, choice of TV programs, contacts with administration, our agenda management, information retrieval etc.

However, to achieve correctly the numerous tasks we shall entrust to them, intelligent agents have to know a lot things about our individual private life. For instance, they need to have information about all our medical records at their disposal to recommend a balanced diet or to prevent medical risks in case of treatment or surgery. But, one may fear that they misuse or disseminate the information they have. How will it be possible to design clever intelligent agents using our personal information when it is needed and only when it is needed? What does it mean that it is needed to use and to divulgate personal information? How one may define and computerize ethical criteria on which intelligent agents could rule their behavior to become our efficient, devoted and faithful servants? That are the questions we want to address in this paper.

Case Based Reasoning vs. Rule Based Reasoning

During the past, there were   many attempts to build computational ethics, i.e. procedures defining ethics for artificial agents or robots (Aaby 2005, Floridi and Sanders 2005, Bringsjord et al. 2006). More precisely, computational ethics models ethical systems by the use of programs and simulates decision procedures with physical information systems, i.e. with computers. Inspired by Asimov’s short story “Runaround” written in 1942 (Asimov 2004), the ethics for artificial agents studies the rules on which robots have to rule their behaviour to be ethically admissible. For instance web agents have to respect privacy; agents in hospitals have to respects patients and their pain etc.

However, one of the difficulties we face when writing rules of behaviour for intelligent agents is that the requirements are numerous and sometimes contradictory. For instance, we want personal robots act as faithful dogs who have to defend and help their master. Simultaneously, we need to protect our privacy by restricting access to personal data. But, we also demand the robot to behave ethically, i.e. to say the truth whenever someone asks them and not to increase information entropy by divulging wrong information. Those three requirements are somehow contradictory, since security of people demands total transparency while personal servants have sometime to lie to protect their master intimacy.   Therefore, agents who pretend to be discreet have to obey to multiple and independent principles that may appear to be contradictory. One solution would be to model ethical reasoning by the use of an analogical inference based on particular cases. But, it would lead to lost the ethical arguments that can only be based on general statements, not on cases.

In other words, when designing ethical intelligent agents, one faces the same difficulties that philosophers faced when they dealt with ethics. One the one hand, a theory – or a science – of ethics requires the use of general statements that can be applied to concrete cases. On the other hand, solving concrete cases leads to many ethical dilemmas since general ethical principles are conflicting each others and inconsistent when applied to specific situations. The debates opposing Platonistic and Aristotelian ethics, or between utilitarian – or empiric – and Kantian ethics, or, more recently, between moral particularism and moral generalism testify the difficulties philosophers have to approach this question.

This paper constitutes an attempt to show that modern logic-based artificial intelligence techniques may contribute to give a new insight, since they were designed to model what is called “common sense reasoning”, which concerns the treatment of similar difficulties, i.e. the management of inconsistent rules on particular cases.

Using Non-Monotonic Logics to Deal with Inconsistent Rules

One of the main difficulties the logic-based artificial intelligence has to deal with is to conciliate the specificity of singular cases with general rules. Depending on the domain of application, it has got different names: “frame problem”, “common sense reasoning”, etc.   To be more precise, let us take an example related to an ethical question: it seems obvious that robots have to serve humans without   being hostile to anyone, whereas when they act as personal servants, they have to fight against their owner's enemies. For the sake of illustration, let us suppose that somebody is attacking you; in such a situation, you would wish your robot defending you, even if it might hurt a human. However, it is preferable that this second rule applies only if the robot owner is not aggressive etc.

Modern logic-based artificial intelligence techniques have been developed to solve this kind of problems within a logical framework. More precisely, the goal of logic-based artificial intelligence techniques is to satisfy rules that don’t lead to contradictions, while being able, in cases of contradictions, to cancel the effects of inconsistent rules.

In the past, many Artificial Intelligence researchers tried to simulate non-monotonic reasoning, i.e. reasoning based on general rules and accepting exceptions. Several formalisms have been developed, for instance, default logic (Reiter 1980), circumscription (McCarthy 1980), non-monotonic logics (McDermott and Doyle 1980), Truth Maintenance Systems, etc. However, most of the mechanical solvers based on those formalisms were very inefficient. Recently, a new efficient and general formalism called Answer Set Programming (ASP) (Baral 2003) has been developed to simulate non-monotonic reasoning. It has been designed to unify previous non-monotonic reasoning formalisms.

For instance, the previous example can be formalized as followed within the ASP formalism:

Øattack(R, H) ¬ robot(R), human_or_robot(H), not ab_attack(R, H).

ab_attack(R, H) ¬ robot(R), human_or_robot(H), owner(R, H1), attack(H, H1), not aggressive(H1).

attack(R, H) ¬ robot(R), human_or_robot(H), ab_attack(R, H).

attack(R, H) or Ø attack(R, H) ¬ .

aggressive(H) ¬ attack(H, H1), not attack(H1, H).

Validating Machine Ethics

Our purpose in this paper is neither to only show how non monotonic logics provide a clear formalization of “machine ethics”, i.e. ethical rules of behavior for intelligent agents, nor to explain how ASP formalisation represents and simulates agent behaviors based on general ethical rules or maxims. Undoubtedly, such a formalisation is useful to design discreet intelligent agents and it would be of practical use. Nevertheless, the main advantage of such a formalisation is that it may lead to go further and to validate of ethical rule systems for robots.

The paper will show that adding to any ethical rule system, a set of general conditions characterizing an harmonious society of men and robots, where everybody can hope to live and to act freely, e.g. stating that men may trust almost anyone without fearing to be betrayed, one may prove that there exists at least one decision that obeys ethical requirements in each situation. In other words, one may define formal conditions – e.g. Local stratification (Baral 2003) – under which a system of ethical rules always leads to at least one decision satisfying the general criteria that characterize an harmonious society of men and machines.

References

Aaby, A. (2005) Computational Ethics, technical report.

Asimov, I. (2004) I, Robot, Spectra, New York, NY.

Baral, C. (2003) Knowledge Representation, Reasoning and Declarative Problem Solving, Cambridge University Press.

Bringsjord, S., Arkoudas, K. and Bello, P. (2006)   Toward a General Logicist Methodology for Engineering Ethically Correct Robots, IEEE Intelligent Systems, vol. 21, number 4, pp. 38–44.

Floridi, L. and Sanders, J. (2004) On the Morality of Artificial Agents, Minds and Machines, 2004, 14.3, pp. 349-379

McCarthy, J. (1980) Circumscription: a form of non-monotonic reasoning. In: Artificial Intelligence, number 13, 27-39, 171-172.

McDermott, D. and Doyle, J. (1980) Non-monotonic logic 1. In: Artificial Intelligence, number 13, pp. 41-72.

Reiter, R. (1980) A logic for default reasoning. In: Artificial Intelligence, number 13, pp. 81-132.