Title: James Bond and Michael Ovitz The Secret Life of Agents
1James Bond and Michael Ovitz The Secret Life of
Agents
- Katia Sycara
- The Robotics Institute
- Carnegie Mellon University
- Pittsburgh, PA 15213
- (412) 268-8825
- katia_at_cmu.edu
- http//www.cs.cmu.edu/sycara
- Project Homepage http//www.cs.cmu.edu/softagent
s
2Team Members CMU
- Liren Chen
- Somesh Jha
- Rande Shern
- Dajun Zeng
- Keith Decker
- Anadeep Pannu
- Vandana Verma
Prasad Chalasani Kostya Domashnev Onn Shehory
3Talk Outline
- Introduction to Agents
- Motivations and advantages of distributed agent
technology - The Retsina Approachª
- Retsina Agent Architecture
- Middle Agents
- Multi-agent interaction protocols (negotiation,
contingent contracting) - Retsina Applications
- Concluding Remarks
- ________________________________
- ª Retsina stands for Reusable Task Structured
Intelligent Networked Agents.
4The (Re)-Emergence of Agents The Marriage of
Two Holy Grails
AI
SE
- Goal-directed
- Adaptive
- Knowledge-based
- Reusability
- Robustness
- Flexibility
Ubiquitous Networked Information Access
Intelligent Software Agents!
5What is an Agent?
- A computational system that
- has goals, sensors and effectors
- is autonomous
- is adaptive
- is long lived
- lives in a networked infrastructure
- interacts with other agents
6Agent vs. Agent
7Next Generation of Agent Technology
- Currently, agent technology is mostly single
agent focusing on information retrieval and
filtering according to user profile - Multi-Agent Systems that interact with humans and
each other - Integrate information management and decision
support - Enable real-time synchronization of the tasks and
actions of humans in teams and organizations - Acquire and disseminate timely and relevant
information - Anticipate and satisfy human information and
problem solving needs - Notify about changes in the environment
- Adapt to user, task and situation
8Motivation for Multi Agent Systems (MAS)
Technology
- Global Information Markets
- Increasingly networked world
- Vast quantities of unorganized information
- Diverse information sources
- Inability for human to manage information access
process---information overloading - Moving from locating documents to making
decisions
9Features of MAS
- Multiple agents connected through communication
networks - Coordination - no agent has sufficient
information or capabilities to solve problem
aloneª - Decentralized control - no master agent
- Decentralized data - no global data storage
- Agent Coupling - balancing computation and
communication - Asynchronous - multiple activities operating in
parallel - _______________________
- ª Agents could be cooperative or self-interested.
10MAS Basic Questions (Bond and Gasser, 88)
- Coherence in coordinated decision making
- Recognition and reconciliation of disparate
viewpoints or conflicting intentions - Graceful performance degradation in face of
missing information and resources - Satisfaction of system-wide criteria (e.g.,
optimality of solution, hard real-time deadlines,
etc.) - How to recognize workload imbalances and
appropriately redistribute activities and
responsibilities among agents - Modeling other agents
- Synthesizing different views and results
11MAS Concepts and Tools
- Distributed Constraint Satisfaction and
Optimization (Yokoo, et al,1991 Sycara, et al.
1991) - Distributed Truth Maintenance and Multi-Agent
Search (Huhns et al. 1991, Ishida 1997) - Organizational Structuring (Lesser and Corkill
1991, Gasser 1993, Decker 1995) - Multiagent Planning (Georgeff 1983, 1984, 1995
Durfee and Lesser 1991, Jennings 1995, Grosz et
al. 1996) - Contracting (Smith 1980, Mueller 1993, Sandholm
1997) - Negotiation (Sycara 1990 Kraus 1991, Zlotkin
1996)
12MAS Concepts and Tools (Contd.)
- Economic and Game Theoretic Techniques
(Rosenschein 1995 Gmytrasiewicz et al. 1991,
Wellman 1993) - Open systems (Hewitt 1991, Gasser 91)
- Multiagent Logics and Ecological Approaches
(Cohen and Levesque 1987 Ferber 1990, Shoham
1993 Huberman et al. 1996) - Social Laws and Norms (Tenenholtz 1991,
Castelfranchi 1993) - Multi-Agent Learning (Sen 1993, Durfee 1994,
Sycara 1996)
13Retsina Approach
- Architecture that includes data- and
knowledge-bases and a distributed collection of
intelligent agentsª - Reusable and composable agent components (agent
editor, agent operating system) - Operates in an open world
- agents, network links, and information sources
appear/disappear - uncertainty
- Dynamic agent team formation on-demand
- ____________________________
- ª K.Sycara, K.Decker, A.Pannu, M.Williamson, and
D.Zeng. Distributed Intelligent Agents. IEEE
Expert, Dec-96
14Retsina Functional Organization
15Main System Issues in MAS
- Single agent architecture
- Retsina agent architecture
- Agent Self-cloning
- Finding other agents
- Middle agents
- Matchmaking and brokering
- Agent interaction protocols
- Negotiation
- Contingent contracting
16Retsina Agent Architecture
- Planning
- hierarchical task network-based formalism
- library of task reduction schemas
- alternative task reductions
- contingent plans, loops
- incremental task reduction, interleaved with
execution - information gathered during execution directs
future planning - Scheduling
- fully expanded leaf nodes executable basic
actions - enabled actions (all parameters and provisions in
place) - adjust periodic tasks with missed deadlines
17- Communication and Coordination
- processes incoming and outgoing messages
- creates new goals/objectives
- determines coordination interactions
- addresses security issues
- Execution Monitoring
- setup execution context (parameters and
provisions) - action monitoring
- deadlines/timeouts
- data collection for decisions (e.g. cloning)
- complete execution (provide results to
appropriate downstream actions)
18Retsina Agent Architecture
19A task Structure (Advertisement Task Structure)
20Reusable Behaviors
- Advertising
- send agent capability model to middle-agent(s)
- shared query behavior for other agents
- Polling for messages
- Answering queries one-shot and periodic
- Monitoring for changes and notification
- Self-cloning
21Self-Cloning Process
- Agents that perceive an overload look for other
agents to pass tasks to (simple model to predict
idle time using learned estimation of task
durations) - When other agents not found
- locate a host with resources for cloning
- create a clone on the host
- partition tasks and transfer to clone
- the old'' agent updates its advertisement the
clone'' agent advertises - When clone is idle -- consider self-extinction,
and shut down if necessary.
22Cloning Experimental Setting
- Number of agents 10 to 20.
- Number of clones allowed up to 10.
- Number of tasks arriving at the system up to
1000. - Task distribution with respect to the required
capabilities for execution normal distribution,
where 10 of the tasks are beyond the
capabilities of the agents. - Agent capabilities an agent can perform up to 20
average tasks simultaneously.
23Task Completion w/wo Cloning
24Middle Agents
- An agent needs to have some task/service
performed. How can it find agents able to perform
that task? - In an open system
- agents generally don't have knowledge of all
other agents - service providers are liable to come and go over
time - A solution middle agents that specialize in
making connections between agentsª - _________________________
- ª K.Decker, K.Sycara, M. Williamson.
Middle-Agents for the Internet. IJCAI-97
25Middle Agent Types
Capabilities Initially Known By
26Matchmaking Agent Yellow Page Services
27Matchmaking in Agent Coordination
- When an agent A advertises its capability, it
Intends toª perform any task that fits the
specification of that capability. - In the Retsina system an agent A advertises a
relational schema SA, i.e., agent A intends to
answer any query on its schema. - If an agent B finds another agent A with a
certain capability through matchmaking, B
believes that agent A can successfully perform
the task. - Matchmaking gives operational semantics to
predicates such as Intend.to, Bel. - _______________________________
- ª Grosz and Kraus 1996
28Performance of Match-made System
29Performance of Brokered System
30Agents in Electronic Commerce
31Adaptive Negotiation (the Bazaar Model)
- Aims at modeling multi-issue negotiation
processesª - Combines the strategic modeling aspects of
game-theoretic models and single agent sequential
decision making models - Supports an open world model
- Addresses heterogeneous multi-agent learning
utilizing the iterative nature of sequential
decision making and the explicit representation
of beliefs about other agents - ______________________________
- ª D.Zeng and K.Sycara. Benefits of Learning in
Negotiation. Proceedings of AAAI-97.
32Utility of Learning Experimental Design
- The set of players N is comprised of one buyer
and one supplier who make alternative proposals. - For simplicity, the range of possible prices is
from 0 to 100 units and this is public
information - The set of possible actions (proposed prices by
either the buyer or the supplier) A equals to 0,
1, 2,, 100 - Reservation prices are private information.
- Each player's utility is linear to the final
price ( a number between 0 and 100) accepted by
both players - Normalized Nash product as joint utility (the
optimal joint utility when full information is
available is 0.25)
33Average Performance of Three Experimental
Configurations in Bazaar
- A non-learning agent makes decisions based solely
on his own reservation price - A learning agents makes decisions based on both
the agent's own and the opponent's reservation
price
34Contingent Contracts and Options
- Most multi-agent systems don't handle uncertainty
effectively - rigid task delegation mechanism (contracts are
binding rather than contingent - no explicit modeling of stochastic events
- no explicit mechanism for controlling agent
performance variability - We are exploring the use of option pricing to
address the above issues
35Evaluation of Contingent vs Binding Contracts
- In the experiments, we only had two kinds of
agents - Interface Agents Accept queries from the user.
- Information Agents Answer queries given by the
Interface agents. - In each cycle a new information agent with a load
randomly distributed between L and 0.9 appears
with probability ?. - When a new information agent comes up, interface
agents have the option to abort the query on the
old information agent and restart it on the new
one. - Interface agents can only switch a bounded number
of queries to the new agent. This is indicated as
Bound in the graphs. - In the experiments the average delay in answering
the queries was measured. This is indicated as
Delay in the graphs.
36Contingent Contracts
L
37Contingent Contracts
Bound
38Contingent Contracting to Handle Unreliability of
Information Sources
- Uncertain waiting time in response to queries
- random network congestion
- uncertain serve congestion/breakdown
39The Query Restart Problem
- Agent A sends query to Agent B.
- Agent B can complete the query in time X, where
- X 1 with probability p.
- X c (c gt 1) with probability 1 - p.
- Expectation EX p (1 - p) c
- If not done by time 1, should agent A abort and
restart, or wait? - Can restarting reduce expectation? The variance?
Both? - Does it help to repeatedly restart k times?
40A Simple Scenario Single restart
- Strategy restart just after time 1, if not done
by then. - Let Xi completion time of i'th query, i 1,2.
- X1, X2 are independent, identically distributed.
- New completion time is Y
- Y
- New expectation
- EY p (1 - p)(1 E X2) (X1, X2 indep.)
- 1 p (1 - p) (1 - p)? c
- If (and only if) c gt 1 1 / p, EY lt X1 !
1 if X1 1, 1 X2 if X1 c.
41A Simple Scenario k Restarts
Number of Restarts k
42Applications
- Visitor Hoster (PLEIADES)
- Satellite Visibility (THALES)
- Portfolio Management (WARREN)
43Characteristics of Retsina
- Open System
- Adaptivity at the agent and organization level
provides robustness - Service-based, economic coordination of agents
- Reusable and extensible domain-independent
computational infrastructure - Integrates information gathering and execution
monitoring with decision making - Framework for addressing uncertainty and
strategic interactions
44Future of Software Agents
- Agent-based software development is an emerging
paradigm - Agent society that parallels human society
- Implication of the emergence of agent society for
human workplaces, institutions, and social
relations - Agent society as a unit of intelligence
- Opportunities and Challenges
- The WEB is a vast knowledge base presenting novel
opportunities for AI - Overall system (human and software agent)
predictability - Security, privacy, trust issues
- Integration of legacy systems
45Overall Issues in Open MAS
- Overall agent organization
- Single agent architecture
- Retsina agent architecture
- Agent Self-cloning
- Finding other agents
- Middle agents
- Matchmaking and brokering
- Agent interaction protocols
- Negotiation
- Contingent contracting
46Overall Issues in Open MAS
- Overall agent organization
- Single agent architecture
- Retsina agent architecture
- Agent Self-cloning
- Finding other agents
- Middle agents
- Matchmaking and brokering
- Agent interaction protocols
- Negotiation
- Contingent contracting
47Overall Issues in Open MAS
- Overall agent organization
- Single agent architecture
- Retsina agent architecture
- Agent Self-cloning
- Finding other agents
- Middle agents
- Matchmaking and brokering
- Agent interaction protocols
- Negotiation
- Contingent contracting