Multi-agent systems (mostly observations on the Electric Elves) Electric Elves Agents revolution: agents have proliferated in human organizations • Personal assistants: Gather information, manage email, shopping… • Control resources: Building temp, software tools, … Next step: Dynamic agent teams facilitate entire organizations • Teams function 24/7 • Agent proxies for humans, helping: – Routine coordination in organizations – Coherent/robust actions to attain organizational goals – Swift reaction to crises • E.g., Coordinate move of personnel, equipment to crisis site • Results applicable to many organizations: military, business,… Illustrative Tasks from USC/ISI Demonstration in Washington, DC: • Rapid team formation: People flying out, support at ISI • Team planning: Travel arrangements, shipping equipment • Team plan monitor/repair: Team member becomes ill, flights delayed, equipment breakdown Hosting visitors at ISI • Team plans/repair: Schedule visit; monitor/reschedule Help at conferences/technical meetings • Team formation/monitor: Arrange meeting with other researchers Facilitate routine organizational activities Current Focus: Elves in One Research Group Mixed 15 agent team: • Agent proxies for 9 researchers (called “Friday”) – Interfaces: PDA/GPS, WAP phones, workstation, fax, speech • Agent proxy for a project assistant • Information agents, schedulers, matchers… Agent proxies run 24/7 • First deployment in a real organization • Help us with real tasks – – – – Coordinate meetings (reschedule if delays, cancel) Decide presenters at research meetings (via auctions) Track people (www.isi.edu/teamcore/info.html) Order our meals Research Challenges • • • • Teamwork and adjustable autonomy in teams Data source verification and reinduction Hybrid logic and topic-based matching Matchmaking for complex agents • • • • • • Dynamic team formation (e.g., via auctions) Human organization norms: authorities, permissions etc. Scale up complexity, number, and heterogeneity Rapid incorporation of new agents Robustness and adaptability of agents Widespread substitutability of agents Focusing on One Research Topic: Adjustable Autonomy in Teams Proxies for users: Teamwork with others, while serving human users Adjustable autonomy: “Dynamically adjust agent’s autonomy” • Autonomous action on behalf of humans reduces burden, but… – Proxies face significant uncertainty, e.g., how hungry? – Errors in autonomous actions may be costly • Reduce autonomy, transfer control to humans in critical situations Teams raise novel challenges for adjust autonomy! • Previous work: Individual agent/user interactions • With teams, an agent must serve the user AND the team E.g., Cannot wait for user input: causes team miscoordination • Pursuing an approach based on C4.5 then Markov Decision Processes Overall E-Elves Architecture, Showing Friday Agents Interacting with Users Elves in Use: Reschedule Meetings Personalize Friday Ordering Dinner “ More & More computers are ordering food,… we need to think about marketing” Subway owner Elves in Use: Wireless Devices WAP Phone PALM VII + GPS Question: presentation • The whole approach to anthropomorphising the assistant process has to be done with care – Probably elves are less loaded than Fridays – Still all sort of room for misinterpretation and setting antisocial norms To act automatically or with request guidance? • Agent (group) task: get all meeting attendees to arrive at same time – But what if one attendee is perceived by his agent as apt to be delayed – User is often better able to determine if the meeting needs to be delayed for him – But potential for mis-coordination while awaiting user response if agent hands the decision over • Agent can – Make an autonomous decision – Transfer control (ask user, and wait) – Change coordination constraints (e.g. delay the meeting a little) Sometimes it goes wrong • Learning defaults by C4.5 (patched with some heuristics) – This won’t always model everything a human would want taken into consideration • Error observed with the elves – Autonomously cancelling a meeting that was desired (e.g. with big boss) (either initially, or after too long of delay from user) – Accepting an invite (to give a presentation) that the user didn’t want – Repeatedly delaying a meeting in small increments (almost 50 times at 5 minutes per) • They’re trying a more sophisticated model – Partially observable Markov decision processes – But the trade-off of autonomy and error in inherent (we’ll come back to this) Privacy and manipulation • The agents contradict ‘little white lies’ – “I was stuck in traffic” “No, you were at the café” – [locked office, lights out] (email): Your agent says you’re in there • Hurt feelings by making importance levels clear – Why are we (e.g. PhD students) given lower priority?! • Allow statistical summaries that embarrass – You’re always 5 minutes late to PhD meetings but on time with staff colleagues! • Manipulation – Stack calendar with dummy meetings, or meetings labelled ‘basketball’ that agent doesn’t know are lower priority, to avoid being selected to give a presentation
© Copyright 2026 Paperzz