## Rational coordination in multi-agent environments (2000)

### Cached

### Download Links

- [deepblue.lib.umich.edu]
- [ranger.uta.edu]
- [msl.cs.uiuc.edu]
- DBLP

### Other Repositories/Bibliography

Venue: | JAAMAS |

Citations: | 19 - 5 self |

### BibTeX

@ARTICLE{Gmytrasiewicz00rationalcoordination,

author = {Piotr J. Gmytrasiewicz and Edmund H. Durfee},

title = {Rational coordination in multi-agent environments},

journal = {JAAMAS},

year = {2000}

}

### Years of Citing Articles

### OpenURL

### Abstract

Abstract. We adopt the decision-theoretic principle of expected utility maximization as a paradigm for designing autonomous rational agents, and present a framework that uses this paradigm to determine the choice of coordinated action. We endow an agent with a specialized representation that captures the agent’s knowledge about the environment and about the other agents, including its knowledge about their states of knowledge, which can include what they know about the other agents, and so on. This reciprocity leads to a recursive nesting of models. Our framework puts forth a representation for the recursive models and, under the assumption that the nesting of models is finite, uses dynamic programming to solve this representation for the agent’s rational choice of action. Using a decision-theoretic approach, our work addresses concerns of agent decision-making about coordinated action in unpredictable situations, without imposing upon agents pre-designed prescriptions, or protocols, about standard rules of interaction. We implemented our method in a number of domains and we show results of coordination among our automated agents, among human-controlled agents, and among our agents coordinating with human-controlled agents. Keywords: coordination; rationality; decision theory; game theory; agent modeling 1.

### Citations

8530 |
Introduction to Algorithms
- Cormen, Leiserson, et al.
- 1990
(Show Context)
Citation Context ...e agents modeled on deeper levels. Thus, to solve the optimization problem on one level requires solutions to subproblems on the lower level. This means that the problem exhibits optimal substructure =-=[18]-=-, and that a solution using dynamic programming can be formulated. The solution traverses the recursive model structure propagating the information bottom-up. The result is an assignment of expected u... |

7052 |
Probabilistic Reasoning in Intelligent Systems
- Pearl
- 1988
(Show Context)
Citation Context ... takes the perspective of the individual interacting agent, with its current subjective state of belief, and coincides with the subjective interpretation of probability theory used in much of AI (see =-=[12, 55, 59]-=- and the references therein). Its distinguishing feature seems best summarized by Myerson ([53, Section 3.6]): The decision-analytic approach to player i’s decision problem is to try to predict the be... |

3697 |
Artificial Intelligence : A Modern Approach
- Russell, Norvig
- 1995
(Show Context)
Citation Context .... In our work, we use the normative decision-theoretic paradigm of rational decision-making under uncertainty, according to which an agent should make decisions so as to maximize its expected utility =-=[17, 20, 23, 32, 40, 68]-=-. Decision theory is applicable to agents interacting with other agents because of uncertainty: The abilities, sensing capabilities, beliefs, goals, preferences, and intentions of other agents clearly... |

1496 |
M.: Reasoning about Knowledge
- Fagin, Halpern, et al.
- 1995
(Show Context)
Citation Context ...roposed it as an extension, decision-theoretic rationality has not yet been included in AOP. The issue of nested knowledge has also been investigated in the area of distributed systems [22] (see also =-=[21]-=-). In [22] Fagin and colleagues present an extensive model-theoretic treatment of nested knowledge which includes a no-information extension (like the no-information model in RMM) to handle the situat... |

1003 |
The contract net protocol: high-level communication and control in a distributed problem solver
- Smith
- 1980
(Show Context)
Citation Context ... communicating according to patterns that the designer deems desirable. Thus, research into coordination techniques has often led to prescriptions for task-sharing protocols, such as the Contract Net =-=[73]-=-, for rules of interaction such as social laws [71], for negotiation conventions [67], and so on. The emphasis in this prior work has been to provide the agents with ready-to-use knowledge that guides... |

778 |
Evolution and the Theory of Games
- Smith, J
- 1982
(Show Context)
Citation Context ...tic equation is the “right” solution without reference to the context in which the quadratic equation has arisen. 24. Binmore [6], as well as others in game theory [14, 15, 42, 43] and related fields =-=[72]-=-, suggest the evolutionary approach to the equilibrating process. The centerpiece of these techniques lies in methods of belief revision, which we also investigated using the RMM framework. Some of ou... |

724 | The knowledge level - Newell - 1982 |

698 |
Agent-oriented programming
- Shoham
- 1993
(Show Context)
Citation Context ...ner deems desirable. Thus, research into coordination techniques has often led to prescriptions for task-sharing protocols, such as the Contract Net [73], for rules of interaction such as social laws =-=[71]-=-, for negotiation conventions [67], and so on. The emphasis in this prior work has been to provide the agents with ready-to-use knowledge that guides their interactions, * This research was supported,... |

602 |
Game Theory: Analysis of Conflict
- Myerson
- 1991
(Show Context)
Citation Context ...y are not directly observable and usually are not known with certainty. In decision theory, expected utility maximization is a theorem that follows from the axioms of probability and utility theories =-=[24, 53]-=-. In other words, if an agent’s beliefs about the uncertain environment conform to the axioms of probability theory, and its preferences obey the axioms of utility theory (see, for example, [68, p. 47... |

492 | Y.: Knowledge and common knowledge in a distributed environment
- Halpern, Moses
- 1990
(Show Context)
Citation Context ...dge if and only if everyone knows p, and everyone knows that everyone knows p, and everyone knows that everyone knows that everyone knows p, and so on ad infinitum. However, in their well-known paper =-=[34]-=-, Halpern and Moses show that, in situations in which agents use realistic communication channels which can lose messages or which have uncertain transmission times, 21 common knowledge is not achieva... |

474 | Games with Incomplete Information played by Bayesian Players
- Harsanyi
- 1967
(Show Context)
Citation Context ...alized, if there happens to be more than one candidate. 24 The definition of the recursive model structure we presented is also closely related to interactive belief systems considered in game theory =-=[3, 37, 52]-=-. Our structures are somewhat more expressive, since they also include the sub-intentional and no-information models. Thus, they are able to express a richer spectrum of the agents’ decision making si... |

473 | Towards flexible teamwork
- Tambe
- 1997
(Show Context)
Citation Context ...ested beliefs, where the beliefs that agents have about how each evaluates game situations can vary [46]. Tambe describes another interesting approach to coordinating agents during team activities in =-=[75]-=-. The applications of game-theoretic techniques to the problem of interactions in multi-agent domains have also received attention in the Distributed AI literature, for example in [65, 66, 67]. This w... |

417 |
The Art and Science of Negotiation
- Raiffa
- 1982
(Show Context)
Citation Context ...tic expected utility maximization with reasoning about other agent(s) that may reason about others, leads to a variant of game theory that has been called a decision-theoretic approach to game theory =-=[3, 7, 41, 62]-=-; we will compare it torational coordination in multi-agent environments 321 traditional game theory further in Section 6. We would also like to remark that there are several ways in which the agent ... |

376 |
Agreeing to disagree
- Aumann
- 1976
(Show Context)
Citation Context ...in Section 3.2, postulating finiteness of knowledge nesting in the recursive model structure. 20 A well-known particular case of infinitely nested knowledge is based on the notion of common knowledge =-=[2]-=-. A proposition, say p, is common knowledge if and only if everyone knows p, and everyone knows that everyone knows p, and everyone knows that everyone knows that everyone knows p, and so on ad infini... |

279 |
Rules of Encounter
- Rosenschein, Zlotkin
- 1994
(Show Context)
Citation Context ...h into coordination techniques has often led to prescriptions for task-sharing protocols, such as the Contract Net [73], for rules of interaction such as social laws [71], for negotiation conventions =-=[67]-=-, and so on. The emphasis in this prior work has been to provide the agents with ready-to-use knowledge that guides their interactions, * This research was supported, in part, by the Department of Ene... |

274 | A market-oriented programming environment and its application to distributed multicommodity flow problems
- Wellman
- 1993
(Show Context)
Citation Context ...ystem as a whole that are desirable by the designer, like stability, fairness and global efficiency. Other work by Koller and Pfeffer [44] on games with imperfect information, Wellman’s WALRAS system =-=[79, 80]-=-, and Sandholm work on coalitions [70] also follow the more traditional lines of equilibrium analysis. 7. Complexity One look at the branching nested representations proposed in this paper is enough t... |

190 |
Individual Decision Making
- Camerer
- 1995
(Show Context)
Citation Context ...ional manner. 2. Some authors have expressed reservations as to the justifiability of these axioms. See the discussions in [50] and the excellent overview of descriptive aspects of decision theory in =-=[10]-=-. 3. The use of expected utility maximization to predict and explain human decision making is widely used in economics. See the overview in [10]. 4. Our implementation uses a KB configured as an ontol... |

183 |
Formulation of Bayesian analysis for games with incomplete information
- Mertens, Zamir
- 1985
(Show Context)
Citation Context ...alized, if there happens to be more than one candidate. 24 The definition of the recursive model structure we presented is also closely related to interactive belief systems considered in game theory =-=[3, 37, 52]-=-. Our structures are somewhat more expressive, since they also include the sub-intentional and no-information models. Thus, they are able to express a richer spectrum of the agents’ decision making si... |

176 | Learning, mutation, and long run equilibria in games
- Kandori, Mailath, et al.
- 1993
(Show Context)
Citation Context ...ide which of the roots of the quadratic equation is the “right” solution without reference to the context in which the quadratic equation has arisen. 24. Binmore [6], as well as others in game theory =-=[14, 15, 42, 43]-=- and related fields [72], suggest the evolutionary approach to the equilibrating process. The centerpiece of these techniques lies in methods of belief revision, which we also investigated using the R... |

167 | Coalitions Among Computationally Bounded Agents
- Sandholm, Lesser
- 1997
(Show Context)
Citation Context ... designer, like stability, fairness and global efficiency. Other work by Koller and Pfeffer [44] on games with imperfect information, Wellman’s WALRAS system [79, 80], and Sandholm work on coalitions =-=[70]-=- also follow the more traditional lines of equilibrium analysis. 7. Complexity One look at the branching nested representations proposed in this paper is enough to suggest that complexity may become a... |

150 |
Probabilistic Reasoning in Expert Systems
- Neapolitan
- 1990
(Show Context)
Citation Context ...nch means that R1 has no information about how it is modeled by R2. Therefore, all of the conjectures that R2 may have about R1’s behavior are possible and, according to the principle of indifference =-=[16, 55]-=-, equally likely. This can be represented as the branch on the first level of the recursive structure splitting into infinite sub-branches, each of which terminates with a different, legal probability... |

143 | Epistemic Conditions for Nash Equilibrium
- Aumann, Brandenburger
- 1995
(Show Context)
Citation Context ...tic expected utility maximization with reasoning about other agent(s) that may reason about others, leads to a variant of game theory that has been called a decision-theoretic approach to game theory =-=[3, 7, 41, 62]-=-; we will compare it torational coordination in multi-agent environments 321 traditional game theory further in Section 6. We would also like to remark that there are several ways in which the agent ... |

130 |
What is intention
- Bratman
- 1990
(Show Context)
Citation Context ...izers, but as mechanisms or simple objects. Apart from game theory we should mention related work in artificial intelligence. In his philosophical investigations into the nature of intentions Bratman =-=[8]-=- distinguishes between mere plans, say as behavioral alternatives, and mental states of agents when they “have a plan in mind” which is relevant for having an intention (see also [1]). Our approach of... |

111 | Preferential semantics for goals
- Doyle, Wellman
- 1991
(Show Context)
Citation Context ... of all of the agents: U � A −→ R, where R is the set of real numbers. Intuitively, a purposeful agent has reason to prefer some actions (that further its purposes in the current situation) to others =-=[78]-=-. Our ability to represent agents’ preferences over actions as payoffs follows from the axioms of utility theory, which postulate that ordinal preferences among actions in the current situation can be... |

109 | Rationality and its roles in reasoning
- Doyle
- 1992
(Show Context)
Citation Context .... In our work, we use the normative decision-theoretic paradigm of rational decision-making under uncertainty, according to which an agent should make decisions so as to maximize its expected utility =-=[17, 20, 23, 32, 40, 68]-=-. Decision theory is applicable to agents interacting with other agents because of uncertainty: The abilities, sensing capabilities, beliefs, goals, preferences, and intentions of other agents clearly... |

107 | Deals among Rational Agents
- Rosenschein, Zlotkin
- 1985
(Show Context)
Citation Context ...m activities in [75]. The applications of game-theoretic techniques to the problem of interactions in multi-agent domains have also received attention in the Distributed AI literature, for example in =-=[65, 66, 67]-=-. This work uses the traditional game-theoretic concept of equilibrium to develop a family of rules of interaction, or protocols, that would guarantee the properties of the system as a whole that are ... |

92 |
Intentional Systems
- Dennett
- 1971
(Show Context)
Citation Context ... as a probability distribution over R2’s possible alternative courses of action. If R1 thinks that R2 attempts to maximize its own expected utility, then R1 can adopt the intentional stance toward R2 =-=[19]-=-, treat R2 as rational, and model R2’s decision-making situation using payoff matrices. R2’s payoff matrix, if it knows about both observation points, arrived at analogously to R1’s matrix above, has ... |

83 | Learning Models of Intelligent Agents
- Carmel, Markovitch
- 1996
(Show Context)
Citation Context ...ts behavior using the description of the state of what is being modeled along with knowledge of its dynamics (like in the qualitative model of a bouncing ball [25], or finite state automata models in =-=[11]-=-). These models can be useful for an agent that can incorporate techniques such as model-based reasoning or qualitative physics to make predictions about the behavior of subintentional 11 entities, re... |

83 | Rosenschein and Gilad Zlotkin. Rules of Encounter: Designing Conventions for Automated Negotiation Among Computers - Jeffrey - 1994 |

80 | defense of probability
- In
- 1985
(Show Context)
Citation Context ... takes the perspective of the individual interacting agent, with its current subjective state of belief, and coincides with the subjective interpretation of probability theory used in much of AI (see =-=[12, 55, 59]-=- and the references therein). Its distinguishing feature seems best summarized by Myerson ([53, Section 3.6]): The decision-analytic approach to player i’s decision problem is to try to predict the be... |

79 |
The Foundations of Expected Utility
- Fishburn
- 1982
(Show Context)
Citation Context ...y are not directly observable and usually are not known with certainty. In decision theory, expected utility maximization is a theorem that follows from the axioms of probability and utility theories =-=[24, 53]-=-. In other words, if an agent’s beliefs about the uncertain environment conform to the axioms of probability theory, and its preferences obey the axioms of utility theory (see, for example, [68, p. 47... |

75 |
Common knowledge
- Geanakoplos
- 1994
(Show Context)
Citation Context ...e than they really do. 22342 gmytrasiewicz and durfee In other related work in game theory, researchers have begun to investigate the assumptions and limitations of the classical equilibrium concept =-=[5, 26, 41, 64, 76]-=-. As we mentioned, our work on RMM follows an alternative approach, proposed in [3, 7, 41, 62], and called a decision-theoretic approach to game theory. Unlike the outside observer’s point of view in ... |

66 |
Modeling Digital Circuits for Troubleshooting
- Hamscher
- 1991
(Show Context)
Citation Context ...rdination in multi-agent environments 329 design stance, which predicts behavior using functionality (such as how the functions of a console controller board’s components lead to its overall behavior =-=[35]-=-), and the physical stance, which predicts behavior using the description of the state of what is being modeled along with knowledge of its dynamics (like in the qualitative model of a bouncing ball [... |

66 | The automated mapping of plans for plan recognition
- Huber, Durfee, et al.
- 1994
(Show Context)
Citation Context ... model, if it has been arrived at without ascribing rationality to the modeled entity. For example, a conjecture as to another’s actions may be derived from plan recognition, from past actions (as in =-=[39]-=-), or from information related by a third agent, and it can be given a probabilistic weight according to the assessment of its faithfulness within the RMM framework. The definition of the recursive mo... |

60 |
Essays on the foundations of game theory
- Binmore
- 1990
(Show Context)
Citation Context ...e than they really do. 22342 gmytrasiewicz and durfee In other related work in game theory, researchers have begun to investigate the assumptions and limitations of the classical equilibrium concept =-=[5, 26, 41, 64, 76]-=-. As we mentioned, our work on RMM follows an alternative approach, proposed in [3, 7, 41, 62], and called a decision-theoretic approach to game theory. Unlike the outside observer’s point of view in ... |

55 | A model-theoretic analysis of knowledge
- Fagin, Halpern, et al.
- 1991
(Show Context)
Citation Context ...le Shoham has proposed it as an extension, decision-theoretic rationality has not yet been included in AOP. The issue of nested knowledge has also been investigated in the area of distributed systems =-=[22]-=- (see also [21]). In [22] Fagin and colleagues present an extensive model-theoretic treatment of nested knowledge which includes a no-information extension (like the no-information model in RMM) to ha... |

52 |
Decision Theory and Artificial Intelligence II: The Hungry Monkey
- Feldman, Sproull
- 1977
(Show Context)
Citation Context .... In our work, we use the normative decision-theoretic paradigm of rational decision-making under uncertainty, according to which an agent should make decisions so as to maximize its expected utility =-=[17, 20, 23, 32, 40, 68]-=-. Decision theory is applicable to agents interacting with other agents because of uncertainty: The abilities, sensing capabilities, beliefs, goals, preferences, and intentions of other agents clearly... |

52 |
Evolution of equilibria in the long run: A general theory and applications
- Kandori, Rob
- 1995
(Show Context)
Citation Context ...ide which of the roots of the quadratic equation is the “right” solution without reference to the context in which the quadratic equation has arisen. 24. Binmore [6], as well as others in game theory =-=[14, 15, 42, 43]-=- and related fields [72], suggest the evolutionary approach to the equilibrating process. The centerpiece of these techniques lies in methods of belief revision, which we also investigated using the R... |

48 | A computational market model for distributed configuration design
- Wellman
- 1995
(Show Context)
Citation Context ...ystem as a whole that are desirable by the designer, like stability, fairness and global efficiency. Other work by Koller and Pfeffer [44] on games with imperfect information, Wellman’s WALRAS system =-=[79, 80]-=-, and Sandholm work on coalitions [70] also follow the more traditional lines of equilibrium analysis. 7. Complexity One look at the branching nested representations proposed in this paper is enough t... |

47 | long run equilibria in games - Learning - 1993 |

44 |
Decision Field Theory: A dynamic cognition approach to decision making
- Busemeyer, Townsend
- 1993
(Show Context)
Citation Context ...rather than an agent proper. 11. That is, not following the decision-theoretic principles of rationality. 12. As we mentioned, we use the expected utility maximization as a descriptive tool. See also =-=[9, 10]-=-. 13. The principle of indifference is applied here to the probability itself. See, for example, the discussion in [16] Section 1.G. 14. We found the method of logic sampling to be an effective approx... |

41 | Learning models of other agents using influence diagrams
- Suryadi, Gmytrasiewicz
- 1999
(Show Context)
Citation Context ...obability of 0.8 to it being rational but not being able to see P2. The remaining no-information model, which includes the possibility of R2’s being irrational, is assigned the probability of 0.1 (in =-=[31, 74]-=- we show how these models and their probabilities can be learned and updated based on the other agent’s observed behavior). Let us note that R2’s best choice of action, in each of the intentional mode... |

40 |
Issues in architectures for planning and execution
- Hanks, Firby
- 1990
(Show Context)
Citation Context ...ent’s actions on the environment and on others’ actions. Other methods include equipping probabilistic or classical planners with multiattribute utility evaluation modules, as in the work reported in =-=[32, 36]-=-, and in our early system called Rational Reasoning System (RRS) [29], which combined hierarchical planning with a utility evaluation to generate the payoff matrices in a nuclear power plant environme... |

37 |
Issues in decision-theoretic planning: Symbolic goals and numeric utilities
- Haddawy, Hanks
- 1990
(Show Context)
Citation Context |

36 |
Knowledge and Equilibrium in Games
- Brandenburger
- 1992
(Show Context)
Citation Context ...tic expected utility maximization with reasoning about other agent(s) that may reason about others, leads to a variant of game theory that has been called a decision-theoretic approach to game theory =-=[3, 7, 41, 62]-=-; we will compare it torational coordination in multi-agent environments 321 traditional game theory further in Section 6. We would also like to remark that there are several ways in which the agent ... |

34 |
Elementary Decision Theory
- Chernoff, Moses
- 1986
(Show Context)
Citation Context ... over actions as payoffs follows from the axioms of utility theory, which postulate that ordinal preferences among actions in the current situation can be represented as cardinal, numeric values (see =-=[13, 20]-=- for details). We represent Ri’s payoff associated with a joint action �a 1 k�����aim�����an l � as uR i a 1 k ···aim ···an . l We now define the recursive model structure of agent Ri� RMSRi , as the ... |

34 |
Spatial and Qualitative Aspects of Reasoning about Motion. AAAI-80
- Forbus
- 1980
(Show Context)
Citation Context ...]), and the physical stance, which predicts behavior using the description of the state of what is being modeled along with knowledge of its dynamics (like in the qualitative model of a bouncing ball =-=[25]-=-, or finite state automata models in [11]). These models can be useful for an agent that can incorporate techniques such as model-based reasoning or qualitative physics to make predictions about the b... |

32 |
Universal subgoaling and chunking: The automatic generation, learning of goal hierarchies
- Laird, Rosenbloom, et al.
- 1986
(Show Context)
Citation Context ... the time and use RMM to determine what the rational response should have been. By storing this as a rule of behavior that can be recalled when appropriate in the future (see related work on chunking =-=[47]-=-), RMM can provide the basis for the accrual of rational heuristics [58]. Another important direction, and an application area, of RMM is studying rational communicative behavior among agents involved... |

31 | Generating and solving imperfect information games
- Koller, Pfeffer
- 1995
(Show Context)
Citation Context ...action, or protocols, that would guarantee the properties of the system as a whole that are desirable by the designer, like stability, fairness and global efficiency. Other work by Koller and Pfeffer =-=[44]-=- on games with imperfect information, Wellman’s WALRAS system [79, 80], and Sandholm work on coalitions [70] also follow the more traditional lines of equilibrium analysis. 7. Complexity One look at t... |

26 |
Subjective Probability and the Theory of Games
- Kadane, Larkey
- 1982
(Show Context)
Citation Context |