Decision-Making Among Bounded Rational Agents

Rational software agents with the BDI reasoning model for Cyber–Physical Systems

This collection began as a conference on Modeling Rational and Moral Agents that combined two themes. First is the problematic place of morality within the received theory of rational choice. Decision theory, game theory, and economics are unfriendly to crucial features of morality, such as commitment to promises. But since morally constrained agents seem to do better than rational agents - say by co-operating in social situations like the Prisoner’s Dilemma - it is difficult to dismiss them as simply irrational. The second theme is the use of modeling techniques. We model rational and moral agents because problems of decision and interaction are so complex that there is much to be learned even from idealized models. The two themes come together in the most obvious feature of the papers: the common use of games, like the Prisoner’s Dilemma (PD), to model social interactions that are problematic for morality and rationality.’

This article argues for a certain picture of the rational formation of conditional intentions, in particular deterrent intentions, that stands in sharp contrast to accounts on which rational agents are often not able to form such intentions because of what these enjoin should their conditions be realized. By considering the case of worthwhile but hard-to-form ‘non-apocalyptic’ deterrent intentions (the threat to leave a cheating partner, say), the article argues that rational agents may be able to form such intentions by first simulating psychological states in which they have successfully formed them and then bootstrapping themselves into actually forming them. The article also discusses certain limits imposed by this model. In particular, given the special nature of ‘apocalyptic’ deterrent intentions (e.g. the ones supposedly involved in nuclear deterrence), there is good reason to think that these must remain inaccessible to fully rational and moral agents.

Today’s complex online applications often require the interaction of multiple (web) services that belong to potentially different business entities. Interoperability is a core element of such an environment, yet not a straightforward one due to the lack of common data semantics. The problem is often approached by means of standardization procedures in a top-down manner with limited adoption in practice. (De facto) standards for semantic interoperability most commonly emerge in a bottom-up approach, i.e., involving the interaction and information exchange among self-interested industrial agents. In this paper, we argue that the emergence of semantic interoperability can be seen as an economic process among rational agents and, although interoperability can be mutually beneficial for the involved parties, it may also be costly and might fail to emerge. As a sample scenario, we consider the emergence of semantic interoperability among rational web service agents in service-oriented architectures (SOAs), and we analyze their individual economic incentives with respect to utility, risk and cost. We model this process as a positive-sum game and study its equilibrium and evolutionary dynamics. According to our analysis, which is also experimentally verified, certain conditions on the communication cost, the cost of technological adaptation, the expected mutual benefit from interoperability, as well as the expected loss from isolation, drive the process.

A standard worry about Hume’s account of justice is that it leaves those who are most vulnerable outside the circumstances of justice. An equally standard reply is that those who are so vulnerable as to fall outside the scope of justice need not thereby fall outside the scope of morality altogether, because on Hume’s account we will often have duties of humanity to treat vulnerable creatures decently. It is not clear that this reply is adequate, for given the apparent priority of justice over natural virtues like those of humanity, it is not clear that duties of humanity provide enough protection for the weak. This paper identifies another problem with Hume’s reply: if those who are extremely vulnerable are nonetheless rational and fully capable of autonomous judgment about how to live, then Hume’s theory still delivers the wrong sorts of protections for them. In particular, it is very plausible to suppose that it would be immoral to engage in paternalistic interference with the decisions of such weak but rational agents about how to live, as long as they are not thereby harming or wronging anyone else. Treating such rational but vulnerable agents paternalistically seems unjust, but Hume’s account cannot vindicate this intuition. Indeed, Hume not only cannot explain why such paternalism is unjust, he seems forced to conclude that we will often be obligated by duties of humanity to engage in such paternalism. For Hume seems committed to the unpalatable conclusion that morality speaks unambiguously in favour of paternalistic interference in such cases as long as we can be reasonably sure that the intended beneficiaries really will be made better off. In this paper, I develop and press this new objection from paternalism against Hume’s account of the circumstances of justice.

Realistic models of decision-making and social interactions, considering the nature of memory and biases, continue to be an area of immense interest. Emotion and mood are a couple of key factors that play a major role in decisions, nature of social interactions, size of the social network, and the level of engagement. Most of the prior work in this direction focused on a single trait, behavior, or bias. However, this work builds an integrated model that considers multiple traits such as loneliness, the drive to interact, the memory, and mood biases in an agent. The agent system comprises of rational, manic, depressed, and bipolar agents. The system is modeled with an interconnected network, and the size of the personal network of each agent is based on its nature. We consider a game of iterated interactions where an agent cooperates based on its past experiences with the other agent. Through simulation, the effects of various biases and comparative performances of agent types is analyzed. Taking the performance of rational agents as the baseline, manic agents do much better, and depressed agents do much worse. The payoffs also exhibit an almost-linear relationship with the extent of mania. It is also observed that agents with stronger memory perform better than those with weaker memory. For rational agents, there is no significant difference between agents with a positive bias and those with a negative bias. Positive bias is linked with higher payoffs in manic and bipolar agents. In depressed agents, negative bias is linked with higher payoffs. In manic agents, an intermediate value of mood dependence offers the highest payoff. But the opposite is seen in depressed agents. In bipolar agents, agents with weak mood dependence perform the best.

Transcendental arguments against skepticism claim that the skeptical argument depends on the falsehood or the unbelievability of the skeptical hypothesis. This chapter argues that the skeptic needs to presuppose the moral or practical rationality of the subject, requiring the existence of an external world with certain features (strongest arguments), the falsehood of the skeptical hypothesis (strong arguments), or the subject’s belief in such a world (weak arguments). The argument starts with rational agency and investigates the sense of moral obligation, moral motives, and virtues that would exist in “vat morality,” arguing that although the skeptic needs to presuppose the rational and moral agency of the subject, the skeptical hypothesis denies or undermines the subject’s agency. The chapter ends by considering whether the skeptic can retreat to Pyrrhonian skepticism to save the skeptical project, concluding that he cannot.

The paper is devoted to mathematical methods of planning in systems consisting of rational agents. An agent is an autonomous object that has sources of information about the environment and influences this environment. A rational agent is an agent who has a goal and uses optimal behavioral strategies to achieve it. It is assumed that there is a utility function, which is defined on the set of possible sequences of actions of the agent and takes values in the set of real numbers. A rational agent acts to maximize the utility function. If rational agents form a system, then they have a common goal and act in an optimal way to achieve it. Agents use the optimal solution of the extreme problem, which corresponds to the goal of the system. The problem of linear programming is considered, in which the number of product sets produced by the system is maximized. To solve the nonlinear problem of optimizing the production plan, the conditional gradient method is used, which at each iteration allows a posteriori estimation of the error of the solution and allows stopping the calculation process after reaching the required accuracy. Since the rational agents that are part of the system can have separate optimality criteria, multi-criteria optimization problems appear. The article considers a humanmachine procedure for solving such problems, which is connected with the conditional gradient method and uses information from the decision-maker (DM) at each iteration. The difficulties of this approach are that the DM is not able to make decisions many times under the condition of a significant number of iterations of the nonlinear programming method. The article proposes to replace OPR with an artificial neural network. Nonlinear and stochastic programming methods are used to find optimal parameters of this network.

The chapter develops the main consequences the present conception of reasons for which people do things has for our status as rational agents. One is that rational agents only need an awareness of things going on around them and the ability to act in response to them, but no faculty of governing themselves. This does away with inner opposition and domination as well as with the restriction of rational agency to the human species. It also does away with activity as the core of agency: we are not masters over what we do for reasons; we do it, sensitive to what we encounter. Another consequence is that we are not, on the strength of being rational agents, subject to moral and other requirements, and that we do not therefore enjoy a special dignity: we are worldly creatures through and through.

This paper deals with modeling mental states of a rational agent, in particular states based on agent?s desires. It shows that the world the agent belongs to forces it to restrict its desires. More precisely, desires of a rational agent are restricted by the constraints that exist in the world and which express what is possible or necessary. Furthermore, if the agent is law-abiding, its desires are restricted by the regulations that are defined in the world and which express what is obligatory, permitted or forbidden. This paper characterizes how desires are restricted depending on the fact that the agent is law-abiding or not. This work considers the general case when the agent orders its own desires according to a preference order. The solution is based on modeling desires, regulations and constraints in an unique formal system which is a logic of conditional preferences.

Considering the popular framing of an artificial intelligence as a rational agent that always seeks to maximise its expected utility, referred to as its goal, one of the features attributed to such rational agents is that they will never select an action which will change their goal. Therefore, if such an agent is to be friendly towards humanity, one argument goes, we must understand how to specify this friendliness in terms of a utility function. Wolfhart Totschnig (Fully Autonomous AI, Science and Engineering Ethics, 2020), argues in contrast that a fully autonomous agent will have the ability to change its utility function and will do so guided by its values. This commentary examines computational accounts of goals, values and decision-making. It rejects the idea that a rational agent will never select an action that changes its goal but also argues that an artificial intelligence is unlikely to be purely rational in terms of always acting to maximise a utility function. It nevertheless also challenges the idea that an agent which does not change its goal cannot be considered fully autonomous. It does agree that values are an important component of decision-making and explores a number of reasons why.

An important result from psycholinguistics (Griffiths & Kalish, 2005) states that no language can be learned iteratively by rational agents in a self-sustaining manner. We show how to modify the learning process slightly in order to achieve self-sustainability. Our work is in two parts. First, we characterize iterated learnability in geometric terms and show how a slight, steady increase in the lengths of the training sessions ensures self-sustainability for any discrete language class. In the second part, we tackle the nondiscrete case and investigate self-sustainability for iterated linear regression. We discuss the implications of our findings to issues of non-equilibrium dynamics in natural algorithms.

Internalism about normative reasons is the view that an agent’s normative reasons depend on her motivational constitution. On the assumption that there are reasons for emotion I argue that (a) externalism about reasons for emotion entails that all rational agents have reasons to be morally motivated and (b) internalism about reasons for emotion is implausible. If the arguments are sound we can conclude that all rational agents have reasons to be morally motivated. Resisting this conclusion requires either justifying internalism about reasons for emotion in a way hitherto unarticulated or giving up on reasons for emotion altogether.

Authors like Keynes, H. Jeffreys and Carnap advocated using a concept of "logical probability". Logical probability had the following properties: (a) it was representable as a function from potential states of full belief (or "evidence") to states of subjective or credal probability judgment. (b) Such functions were alleged to be constrained by principles of probability logic. (c) All rational agents were supposed to be obliged to adopt the standard function that probability logic prescribed. In this essay, it is argued that these three requirements could be satisfied only if probability logic prescribed that credal probability should be numerically determinate. Keynes denied that it should numerically determinate and Carnap abandoned the idea that probability logic could supply a determinate function from states of full belief to numerically determinate credal states that all rational agents ought to adopt. The paper explains that once this is conceded, logical probability ought to be interpreted rather differently than it is customarily is.

Generally, system failures, such as crash failures, Byzantine failures, and so on, are considered as common reasons for the inconsistencies of distributed consensus and have been extensively studied. In fact, strategic manipulations by rational agents are not ignored for reaching consensus in a distributed system. In this paper, we extend the game-theoretic analysis of consensus and design an algorithm of rational uniform consensus with general omission failures under the assumption that processes are controlled by rational agents and prefer consensus. Different from crashing one, agent with omission failures may crash or omit to send or receive messages when it should, which leads to difficulty of detecting faulty agents. By combining the possible failures of agents at the both ends of a link, we convert omission failure model into link state model to make faulty detection possible. Through analyzing message passing mechanism in the distributed system with n agents, among which t agents may commit omission failures, we provide the upper bound on message passing time for reaching consensus on a state among nonfaulty agents and message chain mechanism for validating messages. Then, we prove that our rational uniform consensus is a Nash equilibrium when n > 2t + 1, and failure patterns and initial preferences are blind (an assumption of randomness). Thus, agents have no motivation to deviate the consensus, which could provide interpretable stability for the algorithm in multiagent systems such as distributed energy systems. Our research strengthens the reliability of consensus with omission failures from the perspective of game theory.