Coordination Techniques for Distributed Artificial Intelligence

Coordination, the process by which an agent reasons about its local actions and the (anticipated) actions of others to try and ensure the community acts in a coherent manner, is perhaps the key problem of the discipline of Distributed Artificial Intelligence (DAI). In order to make advances it is important that the theories and principles which guide this central activity are uncovered and analysed in a systematic and rigourous manner. To this end, this paper models agent communities using a distributed goal search formalism, and argues that commitments (pledges to undertake a specific course of action) and conventions (means of monitoring commitments in changing circumstances) are the foundation of coordination in all DAI systems. 1. The Coordination Problem Participation in any social situation should be both simultaneously constraining, in that agents must make a contribution to it, and yet enriching, in that participation provides resources and opportunities which would otherwise be unavailable (Gerson, 1976). Coordination, the process by which an agent reasons about its local actions and the (anticipated) actions of others to try and ensure the community acts in a coherent manner, is the key to achieving this objective. Without coordination the benefits of decentralised problem solving vanish and the community may quickly degenerate into a collection of chaotic, incohesive individuals. In more detail, the objectives of the coordination process are to ensure: that all necessary portions of the overall problem are included in the activities of at least one agent, that agents interact in a manner which permits their activities to be developed and integrated into an overall solution, that team members act in a purposeful and consistent manner, and that all of these objectives are achievable within the available computational and resource limitations (Lesser and Corkill, 1987). Specific examples of coordination activities include supplying timely information to needy agents, ensuring the actions of multiple actors are synchronised and avoiding redundant problem solving. There are three main reasons why the actions of multiple agents need to be coordinated: • because there are dependencies between agents’ actions Interdependence occurs when goals undertaken by individual agents are related either because local decisions made by one agent have an impact on the decisions of other community members (eg when building a house, decisions about the size and location of rooms impacts upon the wiring and plumbing) or because of the possibility of harmful interactions amongst agents (eg two mobile robots may attempt to pass through a narrow exit simultaneously, resulting in a collision, damage to the robots and blockage of the exit). Contribution to Foundations of DAI 2 • because there is a need to meet global constraints Global constraints exist when the solution being developed by a group of agents must satisfy certain conditions if it is to be deemed successful. For instance, a house building team may have a budget of £250,000, a distributed monitoring system may have to react to critical events within 30 seconds and a distributed air traffic control system may have to control the planes with a fixed communication bandwidth. If individual agents acted in isolation and merely tried to optimise their local performance, then such overarching constraints are unlikely to be satisfied. Only through coordinated action will acceptable solutions be developed. • because no one individual has sufficient competence, resources or information to solve the entire problem Many problems cannot be solved by individuals working in isolation because they do not possess the necessary expertise, resources or information. Relevant examples include the tasks of lifting a heavy object, driving in a convoy and playing a symphony. It may be impractical or undesirable to permanently synthesize the necessary components into a single entity because of historical, political, physical or social constraints, therefore temporary alliances through cooperative problem solving may be the only way to proceed. Differing expertise may need to be combined to produce a result outside of the scope of any of the individual constituents (eg in medical diagnosis, knowledge about heart disease, blood disorders and respiratory problems may need to be combined to diagnose a patient’s illness). Different agents may have different resources (eg processing power, memory and communications) which all need to be harnessed to solve a complex problem. Finally, different agents may have different information or viewpoints of a problem (eg in concurrent engineering systems, the same product may be viewed from a design, manufacturing and marketing perspective). Even when individuals can work independently, meaning coordination is not essential, information discovered by one agent can be of sufficient use to another that the two agents can solve the problem more than twice as fast. For example, when searching for a lost object in a large area it is often better, though not essential, to do so as a team. Analysis of this “combinatorial implosion” phenomena (Kornfield and Hewitt, 1981) has resulted in the postulation that cooperative search, when sufficiently large, can display universal characteristics which are independent of the nature of either the individual processes or the particular domain being tackled (Clearwater et al., 1991). If all the agents in the system could have complete knowledge of the goals, actions and interactions of their fellow community members and could also have infinite processing power, it would be possible to know exactly what each agent was doing at present and what it is intending to do in the future. In such instances, it would be possible to avoid conflicting and redundant efforts and systems could be perfectly coordinated (Malone, 1987). However such complete knowledge is infeasible, in any community of reasonable complexity, because bandwidth limitations make it impossible for agents to be constantly informed of all developments. Even in modestly sized communities, a complete analysis to determine the detailed activities of each agent is impractical the computation and communication costs of determining the optimal set and allocation of activities far outweighs the improvement in problem solving performance (Corkill and Lesser, 1986). Contribution to Foundations of DAI 3 As all community members cannot have a complete and accurate perspective of the overall system, the next easiest way of ensuring coherent behaviour is to have one agent with a wider picture. This global controller could then direct the activities of the others, assign agents to tasks and focus problem solving to ensure coherent behaviour. However such an approach is often impractical in realistic applications because even keeping one agent informed of all the actions in the community would swamp the available bandwidth. Also the controller would become a severe communication bottleneck and would render the remaining components unusable if it failed. To produce systems without bottlenecks and which exhibit graceful degradation of performance, most DAI research has concentrated on developing communities in which both control and data are distributed. Distributed control means that individuals have a degree of autonomy in generating new actions and in deciding which tasks to do next. When designing such systems it is important to ensure that agents spend the bulk of their time engaged on solving the domain level problems for which they were built, rather than in communication and coordination activities. To this end, the community should be decomposed into the most modular units possible. However the designer should ensure that these units are of sufficient granularity to warrant the overhead inherent in goal distribution distributing small tasks can prove more expensive than performing them in one place (Durfee et al., 1987). The disadvantage of distributing control and data is that knowledge of the system’s overall state is dispersed throughout the community and each individual has only a partial and imprecise perspective. Thus there is an increased degree of uncertainty about each agent’s actions, meaning that it more difficult to attain coherent global behaviour for example, agents may spread misleading and distracting information, multiple agents may compete for unshareable resources simultaneously, agents may unwittingly undo the results of each others activities and the same actions may be carried out redundantly. Also the dynamics of such systems can become extremely complex, giving rise to nonlinear oscillations and chaos (Huberman and Hogg, 1988). In such cases the coordination process becomes correspondingly more difficult as well as more important1. To develop better and more integrated models of coordination, and hence improve the efficiency and utility of DAI systems, it is necessary to obtain a deeper understanding of the fundamental concepts which underpin agent interactions. The first step in this analysis is to determine the perspective from which coordination should be described. When viewing agents from a purely behaviouristic (external) perspective, it is, in general, impossible to determine whether they have coordinated their actions. Firstly, actions may be incoherent even if the agents tried to coordinate their behaviour. This may occur, for instance, because their models of each other or of the environment are incorrect. For example, robot1 may see robot2 heading for exit2 and, based on this observation and the subsequent deduction that it will use this exit, decide to use exit1. However if robot2 is heading towards exit2 to pick up a particular item and actually intends to use exit1 then there may be incoherent behaviour (both agents attempting to use the same exit) although there was coordination. Secondly, even if there is coherent action, it may not be as a consequence of coordination. For example imagine a group of people are sitting in a park (Searle, 1990). As a result of a sudden downpour all of them run to a tree in the middle of the park because it is the only available source of shelter. This is uncoordinated behaviour because each person has the intention of stopping themselves from becoming wet and 1. Similar experiences have also been noted in organisational science: the greater the task uncertainty, the greater the amount of information which must be processed among decision makers during task execution in order to achieve a given level of performance (Galbraith, 1973). Contribution to Foundations of DAI 4 even if they are aware of what others are doing and what their goals are, it does not affect their action. This contrasts with the situation in which the people are dancers and the choreography calls for them to converge on a common point (the tree). In this case the individuals are performing exactly the same actions as before, but it is coordinated behaviour because they each have the aim of meeting at the central point as a consequence of the overall aim of executing the dance. For these two reasons, the coordination process is best studied by examining the internal structure of the individual agents (i.e. the agents’ beliefs, desires, preferences, intentions, and so on). Having decided upon a perspective, the next decisions concern the model that will be used to describe the problem and the structures that will be used to describe the agents. Here a distributed goal search formalism is used to characterise DAI systems (section 2) and the key agent structures are commitment and convention (section three). This model of coordination is founded upon the “Centrality of Commitments and Conventions Hypothesis” which states that: all coordination mechanisms can ultimately be reduced to commitments and their associated (social) conventions. Commitments are viewed as pledges to undertake a specified course of action, while conventions provide a means of monitoring commitments in changing circumstances. The former provide a degree of predictability so that agents can take the (future) activities of others into consideration when dealing with inter-agent dependencies, global constraints or resource utilization conflicts. The latter provide the flexibility which cooperating agents need if they are to cope with being situated in dynamic environments. To operate effectively when the external world and their own beliefs are constantly changing, agents must possess a mechanism for evaluating whether existing commitments are still valid. Conventions provide this mechanism: defining the conditions under which commitments should be reassessed and specifying the associated actions which should be undertaken in such situations. Finally, section four investigates three prominent coordination techniques (organisational structuring, meta-level information exchange and multi-agent planning) and shows how they can all be reformulated in terms of commitments and conventions thus providing further evidence for the main claim of this paper. 2. Modelling Distributed AI Systems as a Distributed Goal Search Problem Several authors have recently characterised DAI as a form of distributed goal search with multiple loci of control (Durfee and Montgomery, 1991; Gasser, 1992; Jennings, 1993; Lesser, 1991). Adopting Lesser’s basic formalism, the actions of Agent1 and Agent2 in solving goals G0 and G 2 0 respectively can be expressed as a classical AND/OR goal structure search 2 (figure 1). The classical graph structure has been augmented to include a representation of the interdependencies between the goals and to indicate the resources needed to solve the primitive goals (leaf nodes). Interdependencies can exist between high level sibling goals, such as G1 and G 1 2, or they can be more distant in the goal structure (eg between G 1 1,1 and G 2 p,2). In the latter case, G1 and G 2 p become interacting goals if G 1 1,1 is used to solve G 1 1. Indirect dependencies exist between goals through shared resources (eg Gm,1,2 and G 2 p,2,2 through resource dj). Resource dependencies can be removed simply by providing more of the resource in question; dependencies between goals, on the other hand, cannot be circumvented as they are a logical consequence of the community’s environment. In all other aspects, the two types of dependency are identical. 2. Figure 1 represents a typical multi-agent situation in which each individual has its own goals, but it must interact with others to achieve them. In contrast, a distributed problem solving system would have a single root node corresponding to the common objective. Contribution to Foundations of DAI 5 Interdependencies can be classified along two orthogonal dimensions: whether they are weak or strong, and whether they are uni-directional or bi-directional. Strong dependencies must be satisfied if the dependent goal is to succeed; weak dependencies facilitate or constrain problem solving but need not be fulfilled for the dependent goal to succeed. An example of a strong dependency is where the output of a goal (G) is a mandatory input (I) for the dependent goal (DG) and where G is the only source of I in the community. A weak dependency exists if there is more than one source for I or if I is an optional input for DG. A uni-directional dependency (written G1,1 → Gp,2) means that agent2’s goal Gp,2 is dependent (either strongly or weakly) on agent1’s goal G1,1, but G 1 1,1 is unaffected by G 2 p,2; with bi-directional dependencies (written Gm,1 ↔ Gm,2) the goals of both agents are affected. The provision of information I by goal G for DG is an example of a uni-directional dependency (G → DG); a bi-directional dependence occurs, for example, when two goals need to be performed simultaneously. The nature of the inter-agent dependencies is the critical determinant of the type of coordination which will take place. For example, if Agent1 knows that G 2 p,2,2 requires resource dj before it can start (strong dependency, uni-directional), then it may decide to execute Gm,1,2 (to produce the necessary resource) before G 1 m,1,1 if there is no other information distinguishing between these two alternatives. Secondly, the relationship between Gm,1 and G 2 m,2 may stipulate that both actions need to be performed simultaneously (strong dependency, bi-directional) in which case the two agents need to reach an agreement about the respective execution times. Finally, if Agent1 chose G 1 1,1 as a means of satisfying G 1 1 the result of this task may provide valuable information (weak dependency, uni-directional) which Agent2 could use when solving G 2 p,2 (eg it may provide a partial result which enables G 2 p,2 to AND Gp,1,3 (G 2 p,1,4) G 2 p,2,2 G0 G1 G 1 2 ......... G 1 k G 1,2 m G 2 p ......... G 2 t G0