DOC PREVIEW
Team Oriented Programming and Proxy Agents: The Next Generation

This preview shows page 1-2-3-4 out of 12 pages.

Save
View full document
View full document
Premium Document
Do you want full access? Go Premium and unlock all 12 pages.
Access to all documents
Download any document
Ad free experience
View full document
Premium Document
Do you want full access? Go Premium and unlock all 12 pages.
Access to all documents
Download any document
Ad free experience
View full document
Premium Document
Do you want full access? Go Premium and unlock all 12 pages.
Access to all documents
Download any document
Ad free experience
View full document
Premium Document
Do you want full access? Go Premium and unlock all 12 pages.
Access to all documents
Download any document
Ad free experience
Premium Document
Do you want full access? Go Premium and unlock all 12 pages.
Access to all documents
Download any document
Ad free experience

Unformatted text preview:

Team Oriented Programming and Proxy Agents: The NextGenerationPaul Scerri, David V. Pynadath, Nathan Schurr, Alessandro Farinelli,Sudeep Gandhe and Milind TambeComputer Science DepartmentUniversity of Southern California{scerri,pynadath,schurr,farinelli,gandhe,tambe}@usc.eduApril 16, 2003AbstractCoordination between large teams of highly hetero-geneous entities will change the way complex goalsare pursued in real world environments. One ap-proach to achieving the required coordination in suchteams is to give each team member a proxy that as-sumes routine coordination activities on behalf of itsteam member. Despite that approach’s success, aswe attempt to apply this first generation of proxy ar-chitecture to larger teams in more challenging envi-ronments some limitations become clear. In this pa-per, we present initial efforts on the next generationof Team Oriented Programming (TOP) and proxyarchiecture, called Machinetta. Machinetta aims toovercome the limitations of the previous generationof proxies and allow effective coordination betweenvery large teams of highly heterogeneous agents. Wedescribe the principles underlying the design of theMachinetta proxies and present initial results fromtwo domains.1 IntroductionDespite their successes, the first generation proxy ar-chitectures suffer from three key limitations whenhandling: scale, dynamism, and effective integrationof humans in agent teamwork. First, in small-scaleteams, agents can be allocated to roles by hand priorto starting up team activities; although limited re-allocation can occur at run-time. Unfortunately, inlarge-scale teams, off-line allocation of agents to rolesby hand is difficult. Second, a high level of dynamismin the environment requires that agents’ role allo-cation and reallocation strategy must often be inte-grated into one unified fluid algorithm, rather thanas two separate phases (allocation vs reallocation) asseen in existing architectures. Furthermore, agentsmust consider role reallocation not only under catas-trophic failures (as was done previously), but mustbe willing to give up current roles to take up newprecious opportunities. Third, as we build increas-ingly heterogeneous teams, and particularly includehumans in the loop, we must enable the proxies totap into human expertise in coordination in key sit-uations. Previous research on teamwork has allowedagents and humans to work together, but the humanparticipation was limited to domain level activities.In this work, we propose to go beyond, to enablehumans to provide their valuable expertise in teamcoordination activities as well.In this paper, we present initial efforts on the nextgeneration of TOP and proxy archiecture, called Ma-chinetta. Machinetta aims to overcome the limita-tions of the previous generation of proxy and alloweffective coordination between very large teams ofhighly heterogeneous agents. To achieve this Ma-chinetta embodies several new design principles, fo-1cused on overcoming the limitations of the previ-ous proxies. To address the first two limitationsdiscussed above, Machinetta has a fluid, integratedrole allocation and reallocation algorithm. Withinthis algorithm, agents attempt to continually allocateand reallocate themselves to new tasks. When newtasks/opportunities arise, or when agents’ capabili-ties decline substantially, agents reconsider their cur-rent commitments to roles; thus, agents may changeroles even without catastrophic failures. This new in-tegrated algorithm enables a much more flexible re-sponse to dynamic environments. However, the an-swer is not simply to replace current coordination al-gorithms with new ones. If a big enough team is putinto a complex enough environment there are, despiteour best efforts, bound to be situations where any co-ordination algorithms perform very poorly or breaksaltogether. A key idea in Machinetta is to acknowl-edge that such problems are going to occur and buildin mechanisms for meta-reasoning to handle those sit-uations. This is achieved by making as much of thecooridnation process as possible explicit, thus makingit easier to monitor the coordination and understandwhen problems occur. For example, we use a roleallocation algorithm[11] that represents each role tobe allocated as an explicit role. If the role of allocat-ing the role goes unachieved for some period of time,i.e., because the standard role allocation algorithmdoes not suceed in allocating it, the team can detectthis situation and recursively invoke meta-reasoningabout ”role allocation role”.With respect to involving humans in coordina-tion (and not just in domain-level tasks), the meta-reasoning capability provides a helpful mechanism.In particular, when meta-reasoning about coordina-tion, agents can appeal to human input. However,humans could provide input that may not necessarilybe in agreement with choices made by the coordina-tion algorithm. Thus, given the possibility of sucharbitrary changes by humans to coordination algo-rithms, the algorithms must be robust to decisionsthat are “wrong” according the algorithm. For ex-ample, a human may arbitrarily (so far as the prox-ies are concerned) decide to terminate a plan and theproxies must implement this decision.The final change in direction for the new genera-tion of proxy is the properties that we aim to provefor the key algorithms. With relatively small teams,establishing properties such as optimality is impor-tant. However, typically proofs of such algorithmproperties rely on assumptions such as the underly-ing situation not changing while the algorithm is ex-ecuting. While such assumptions are very reasonablefor small teams, they are not so interesting for verylarge teams where the assumptions will never be met.The critical point is that large enough teams in com-plex enough environemts will be in a constant state ofchange. For example, in a large team for disaster re-covery in a large city, some team member will alwaysbe completing, abandoning or beginning a task. Theinherent, continuous dynamics makes other algorith-mic properties interesting. For example, the ”stabil-ity” of the system – will one team member’s failureto complete a role lead to many role reallocations orwill the effects be limited. Another interesting prop-erty would be to show that certain events will neverhappen or happened only with a very low probabil-ity. For example, we may be able to prove that someteam member will always eventually accept a role, ifits


Team Oriented Programming and Proxy Agents: The Next Generation

Download Team Oriented Programming and Proxy Agents: The Next Generation
Our administrator received your request to download this document. We will send you the file to your email shortly.
Loading Unlocking...
Login

Join to view Team Oriented Programming and Proxy Agents: The Next Generation and access 3M+ class-specific study document.

or
We will never post anything without your permission.
Don't have an account?
Sign Up

Join to view Team Oriented Programming and Proxy Agents: The Next Generation 2 2 and access 3M+ class-specific study document.

or

By creating an account you agree to our Privacy Policy and Terms Of Use

Already a member?