Coordinating Sociall Constraints and Conversatio

We explore the view that coordinated behavior is explained by the social constraints that agents in organizations are subject to. In this framework, agents adopt those goals that are requested by their obligations, knowing that not fulfilling obligations induces a price to pay or a loss of utility. Based on this idea we build a coordination system where we represent the organization, the roles played by agents, the obligations imposed among roles, the goals and the plans that agents may adopt. Once a goal adopted, a special brand of plans, called conversation plans, are available to the agents for effectively carrying out coordinated action, Conversation plans explicitly represent interactions by message exchange and their actions are dynamically reordered using the theory of Markov Decision Processes to ensure the optimization of various criteria. The framework is applied to model supply chains of distributed enterprises. Introduction and Motivation To build autonomous agents that work coordinately in a dynamically changing world we have to understand two basic things. The first is how an agent chooses a particular course of action and how its choices change in face of events happening in the world. The second is how agents execute coordinated actions. In this paper we present a framework that answers these questions by construing agents as rational decision makers that exist within organizations. Organizations are systems that constrain the actions of member agents by imposing mutual obligations and interdictions. The association of obligations and interdictions is mediated by the roles agents play in the organization. For example, when an agent joins a software production organization in the system administrator role, he becomes part of a specific constraining web of mutual obligations, interdictions and permissions social constraints of laws that link him as a system administrator to Copyright 01997, American Association for Artificial Intelligence (www.aaai.org). All rights reserved. 16 AGENTS developers, managers and every other role and member of the organization. Not fulfilling an obligation or interdiction is sanctioned by paying a cost or by a loss of utility, which allows an agent to apply rational decision making when choosing what to do. Social laws are objective forces that provide the ultimate motivation for coordinated action at the organization level and to a large extent determine the mental states at the individual agent level. Agents “desire” and set as “goals” the things that are requested by their current obligations, knowing that otherwise there will be a cost to pay. However, current models of collective behavior largely ignore this aspect, trying to explain coordination solely from the perspective of socially unconstrained individuals. For this reason they often impose restrictive conditions that limit the generality of the models. For example, the Cohen-Levesque account of teamwork (Cohen & Levesque 91) requires team members to have the same mutual goal. But this is not true in organizations, for example it is normal for supervisors to decompose and schedule work and assign team members different goals which they can carry out coordinately often without being even aware of each other’s goals. Similarly, dropping off a team, according to the Cohen-Levesque model, requires an agent to make his goal to convince every member of the team that, e.g., the common motivation for the joint action has disappeared. Since common goals or motivations do not necessarily exist, this is also not true in general. Imagine an agent having to convince everybody else that he’s got a better job elsewhere before he can leave the organization! To initiate a more realistic investigation of such social constraints and of their role in achieving coordinated behavior, we have built an agent coordination framework whose building blocks include the entities social constraints are made of: agents, organizations, roles, obligations, interdictions, permissions, goals, constraints, plans, etc. In this framework agents determine what obligations and interdictions currently From: AAAI-97 Proceedings. Copyright © 1997, AAAI (www.aaai.org). All rights reserved. apply and on this basis decide on their goals, even when these obligations are in contradiction. Once an agent has chosen a goal, it selects a plan to carry it out. Plans are described in a planning formalism that explicitly represents interactions with other agents by means of structured conversations involving message exchanges (hence the name of conversation plans). Plan actions are dynamically ordered using a Markov Decision Process method to maximize certain kinds of rewards. Details follow. The Vocabulary for Social Constraints We start by briefly introducing our application domain, the integration of supply chains of manufacturing enterprises. A supply chain is a globally extended network of suppliers, factories, warehouses, distribution centers and retailers through which raw materials are acquired, transformed into products, delivered to customers, serviced and enhanced. The key to the efficient operation of such a system is the tight coordination among components. But the dynamics of the enterprise and of the world market make this difficult: customers change or cancel orders, materials do not arrive on time, production facilities fail, workers are ill, etc. causing deviations from plan. Our goal is thus to achieve coordinated behavior in dynamic systems of this kind by applying agent coordination technologies. At the highest level, the above defines an organization where agents play various roles. An organization consists of a set of roles filled by a number of agents. In the example below, customer, coordinator etc. are roles filled respectively by agents Customer, Logistics, etc. (def -organization SC1 :roles ((customer Customer) (coordinator Logistics) (assembly-plant Plant11 (painting-plant Plant21 (transportation Transpl))) An agent can be a member of one or more organizations and in each of them it can play one or more roles. An agent is aware of the existence of some of the other agents, but not necessarily of all of them. Each agent has its local store of beliefs (taken as a data base rather than mental states). A role describes a major function together with the obligations, interdictions and permissions attached to it. Roles can be organized hierarchically (assembly-plant and painting-plant would be both manufacturing roles) and subsets of them may be declared as disjoint in that the same agent can not perform them (for example manufacturing and transport). For each role there may be a minimum and a maximum number of agents that can perform it (e.g. minimum and maximum 1 president). Obligation, Interdiction, Permission. An agent al in role r 1 has an obligation towards an agent a2 in role r2 for achieving a goal G according to some constraint C iff the non-performance by ai of the required actions allows a2 to apply a sanction to al. Agent a2 (who has authority) is not necessarily the beneficiary of executing G by the obliged agent (you may be obliged to your manager for helping a colleague), and one may be obliged to oneself (e.g. for the education of one’s children). In our language we provide a construct for defining obligations generically. The generic obligation exists between two agents in specified roles, whenever a given condition applies. The obligation requires the obliged agent to achieve a goal under given constraints. For example: (def-obligation Reply-to-inquiry : obliged coordinator :authority customer :condition (received-inquiry :from (agent-playing customer) :by (agent-playing coordinator)) :goal Reply-to-sender :enforced (max-reply-time 5)) The above requires the coordinator agent (the : obliged party) to reply to an inquiry from the customer (the : authority party) in at most five units of time (a constraint on the goal Reply-to-sender). This generic obligation becomes active when its condition is satisfied, in this case when the coordinator receives an inquiry from the customer. When this happens, an actual obligation is created linking the coordinator to the customer and applying to the actual inquiry received (if many inquiries are received, as many actual obligations are created). In exactly the same manner, our language defines generic and actual interdictions (the performance of the goal is sanctioned) and permissions (neither the performance nor non-performance are sanctioned). We represent permissions explicitly because we do not assume everything not explicitly obliged or forbidden to be permitted. Agents may choose their goals from their explicit permissions, or requests that can not be proven as obligatory may be served as permissions. Finally, the obligations, interdictions and permissions (short OIP-s) of a role are inherited by sub-roles. Goal. Like OIP-s, goals are described in both a generic and an actual form. In the generic form, a goal specification comprises a list of super-goals, lists of goals incompatible with it, any disjunctive coverage (when a number of subgoals are all the possible sub-