Presentation is loading. Please wait.

Presentation is loading. Please wait.

Electric Elves: Towards Multi-Agents-Facilitated Human Organizations

Similar presentations


Presentation on theme: "Electric Elves: Towards Multi-Agents-Facilitated Human Organizations"— Presentation transcript:

1 Electric Elves: Towards Multi-Agents-Facilitated Human Organizations
Milind Tambe Craig Knoblock Yolanda Gil Hans Chalupsky David Pynadath Tom Russ Funded by DARPA COABS

2 Electric Elves Agents revolution: agents have proliferated in human organizations Personal assistants: Gather information, manage , 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,…

3 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

4 Potential Elves Applications
Military crisis response Teaming/coordination of military units/equipment NIMA & MITRE MITRE has suggested deployment of Elves, for transitions to NIMA, others… Fits well with NIMA scenario requirements IDA “PSYOP” analysis IDA proposed Elves for coordinating analysts & tools Interest from NASA (e.g., space station)

5 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 ( Order our meals

6 Research Challenges Teamwork and adjustable autonomy in teams [Teamcore] Data source verification and reinduction [Ariadne] Hybrid logic and topic-based matching [Loom] Matchmaking for complex agents [Expect] 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

7 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 Markov Decision Processes

8 Elves Architecture Overview: Agents Running on Agent Grid
Flight Monitor Schedule Extractor Flight info Interest Matcher Meet maker Scheduler agent Capability Matcher Teamcore Interest Finder Ariadne Paper Titles abstract Research Publication agent Loom Expect

9 Elves Architecture Overview: Teamcore Proxies on Agent Grid
Bids for Auctions: Tambe: Capable, Willing Pynadath: Capable, …. AGENT GRID Flight Monitor Teamcore proxy Teamcore proxy Teamcore proxy Schedule Extractor Flight info Interest Matcher Teamcore proxy Teamcore proxy Teamcore proxy Meet maker Scheduler agent Capability Matcher Teamcore Interest Finder Ariadne Paper Titles abstract Research Publication agent Loom Expect

10 Elves in Use: Reschedule Meetings
Personalize

11 Friday Ordering Dinner
“ More & More computers are ordering food,… we need to think about marketing” Subway owner

12 Elves in Use: Wireless Devices
WAP Phone PALM VII + GPS

13 Elves Results Teamcore proxies 24/7, 6 month deployment in a real organization Use of Elves virtually eliminated in group about delays, … Use of Elves also eliminted about scheduling talks Elves used to order 54 meals, 1.4 people served on avg per order Num meetgs Total reschedule Auto reschedule Human reschedule 1128 285 235 50 # meet Auto decisions Max bids Avg bids Auto winner bids 8 6 9 6.2 <capable,willing> <not capable,willing>

14 Electric Elves: 24/7 Run in a Real Organization

15 USA Today Article

16 Summary Unique experiment: 24/7 agent teams deployed in a real organization Help in daily organization activities, relevant to other organizations Some lessons: Real-world application critical: Provided real data & challenges Agent grid: Allows interoperation despite different platforms/language Adaptability: Critical to dealing with a changing environment Working together: Elves collaboration accomplished more

17 Future Work Research: Scale-up to ~100 agents (ISI’s Intelligent Systems Division) Versatile interfaces with humans for tasking Human organizations and norms Automated team self-repair and learning Privacy Security Future collaborations: NIMA/MITRE IDA for PSYOPS NASA…

18 Next Steps: Large-scale Virtual Organizations
Virtual community dedicated to a common goal Brings together specialized skills/resources; No one organization possesses required skill Organizations formed from elements from multiple organizations People, agents, software/computational resources, equipment,.. . Apply to rescue in large-scale urban disasters, e.g., earthquakes Collaborate with international efforts on Robotic Rescue (e.g., simulate Kobe quake) Collaborate with SCEC Bring together agent & computational grid

19 THANK YOU

20 Electric Elves Novelty: Teamwork in Teamcore Proxies
Teamcore proxies use a general teamwork model “STEAM” Domain independent, reusable rules of teamwork Explicitly outline team member’s responsibility in teamwork Used in helicopter combat simulations, RoboCup Soccer, … Now in Teamcore proxies to team up heterogeneous agents Teamcore: Reduce development cost in building teams Teams tasked via high-level team-oriented programs (TOP) TOP specifies team plans to be executed, but not coordination Teamcore proxies automatically execute required coordination Maintain coherence in team, but with selective communication Reorganize roles via substitution if a role fails

21 STEAM Rules in Teamcore: Example
If Agent A1’s private state contains a fact F AND Fact F matches an achievement condition AC of a team plan Fact F is not currently mutually believed… THEN Create possible communicative goal to communicate fact F to team 300 rules: Joint Intentions (Cohen/Levesque) & SharedPlans (Grosz) Actual implementation, with modifications to theories proposed Continue to push development of general teamwork theories

22 Novel Issues in Adjustable Autonomy
Novel challenge: Serve individual & team simultaneously Challenge 1: Transferring autonomy control difficult with teams If agent transfers control & waits for user input: miscoordinate Unlike individual user-agent systems, cant wait for input If no transfer control, no user input & act uniformed, then fail Challenge 2: Ensuring desirable team-level decisions Despite appropriate individual-level actions, team-level failures Challenge 3: “Safe learning” for guaranteed behavior

23 Adjustable Autonomy (Cont)
Initially: Use C4.5 decision trees (Inspired by CAP [Mitchell, 94]) Learn user preferences, e.g., if ISI, 5 min to meet, delay 15 Learn whether to transfer autonomy control to user or not Despite initial success some dramatic failures Cancelled meeting with division director, automatically volunteered for presentation, meeting delayed to 8 pm or 9 pm… Now: Autonomy policies using markov decision processes (MDP) Explicitly reason with cost & uncertainty in transferring control Not rigidly commit to decisions of transferring autonomy Don’t take incorrect action if high risk; change coordination E.g., Delay a meeting to get more time for user input

24 Elves in Use: An MDP Autonomy Policy in Use
Visitor at ISI, picked up at 9:00 AM from LAX for 9:30 meeting (9/11) 09/10/00 21:30: Prefer wait in {Time: 'Early', Location: 'not ISI'} 09/11/00 09:15: Prefer wait in {Time: -15, Location: 'not ISI’} 09/11/00 09:25: Prefer delay_5 in {Time: -5, Location: 'not ISI‘} Time=9:23:16, Latitude=N ,Longitude=W 09/11/00 09:25: Asking... 09/11/00 09:30: Prefer delay_15 in {Time: 0, Location: 'not ISI‘} Time=9:27:20, Latitude=N ,Longitude=W 09/11/00 09:25: Asking... 09/11/00 09:35: Prefer delay_15 in {Time: 5, Location: 'not ISI'} Time=9:33:11, Latitude=N ,Longitude=W … 09/11/00 09:35: Acting to delay meeting 15 minutes

25 Electric Elves in Use: General Observations
Multi-agent 24/7 deployment in a real organization Assists us in our daily activities No s have been sent within group about delays, cancels,… No s about scheduling talks at research meeting recently Web page locator helps quickly locate individuals Mobile devices keep us informed about schedule, meet status… Friday not passive: gets us to the meeting on time Ordering lunch ahead is time saver, and a nice surprise

26 Electric Elves in Use: Research Talks Auction
Does not require all bids to be in; the highest bid may not be 1,1 Meeting Date Num of bids Highest Bid Autonomous Decision? 7/6/2000 7 Scerri <1,1> No 7/20/2000 8 Yes 7/27/2000 Kulkarni <0,1> 8/03/2000 Nair <1,1> 8/31/2000 4 Tambe <1,1> 9/21/2000 3 Visitor <-,-> 10/31/2000 Tambe<1,1>

27 Electric Elves in Use: Meeting Data

28 Electric Elves: Some Experiments
If likelihood of user response higher: Agents will ask more If cost of asking users higher: Agents will ask less

29 Focusing in on a Month (June)
Number of messages June, 2000

30 Synergies Teamcore roles assigned using Phosphorus
Teamcore uses capabilities matcher to determine speakers for meetings Teamcore proxies access on-line sources through Ariadne wrappers Agents make decisions based on people’s calendars, flight schedules, etc. Ariadne’s wrappers and IR used in PowerLoom interest matcher Matcher’s KB identifies people’s interests from publications

31 Agent Organization is Grounded in the Real World
On-line information sources Flight schedules, restaurants, etc. Calendar information “Sensors” and “actuators” GPS, “finger”, etc. Postpone/cancel meetings, faxes, etc. Interaction with people Communicate through personal portable devices Support routine tasks

32 Elves in Use: Auctions PAUL SCERRI WILL BE THE PRESENTER FOR TEAMCORE Meeting on 7/20/2000 If presenter cannot attend the meeting, then substitute


Download ppt "Electric Elves: Towards Multi-Agents-Facilitated Human Organizations"

Similar presentations


Ads by Google