Transcript Agent Tech
An Introduction to Agent
Technologies
Peter Wurman, NCSU
Yelena Yesha, UMBC
Olga Streltchenko, UMBC
1
Presentation Overview
Working definition of an agent
Agent characteristics and properties
Agent societies
Examples
2
Working definition of an
agent
“Agents are active, persistent (software)
components that perceive, reason, act, and
communicate”
Huhns and Singh, 1998
“An agent is an entity whose state is viewed as
consisting of mental components such as
beliefs, capabilities, choices, and commitments.
[sic] In this view, therefore, agenthood is in the
mind of the programmer.”
Shoham, 1993
3
Agent Program
Inputs = observations
Observations: states of the agent’s domain or
environment
Outputs = actions
Actions: Speak, Search, Move, Bid
( o1, o2, … )
( a 1 , a2 , … )
Agent
4
Agents Environments
An agent must have a model of its domain
and a model of other agents that it
communicates with.
Properties of agents’ environment:
Observable
Dynamic
Discrete
5
Basic Characteristics
Delegation abilities: The owner or user of
an agent delegates a task to the agent
and the agent autonomously performs the
task on behalf of the user.
An agent can decompose and/or delegate
the task to other agents;
Once the task is complete the agent may
need to report to the user/agent issuing the
task.
6
Basic Characteristics
(cont’d)
Agent communication languages and protocols:
information exchange with other agents
establishes a need for expressive communication
and negotiation language.
KQML (Knowledge Query and Manipulation
Language);
Used to allow information agents to assert interests in
information services, advertise their own services, and
explicitly delegate tasks and requests for assistance from
other agents.
Can be used for developing a variety of inter-agent
communication protocols that enable information agents to
collectively cooperate.
7
Basic Characteristics
(cont’d)
Self-representation abilities: the ability to
express business and system aspects of
its functionality, combine them into an
application or implementation.
Self-describing, dynamic reconfigurable
agents;
Facilitate composition (specification and
implementation) of large-scale (distributed)
applications.
8
Agent Mental State (BDI)
Beliefs–knowledge about the world and
the effects the agent’s actions have on the
world.
Desires–preferences over possible states
of the world (goals).
Intentions–internal commitments made to
achieve certain world states.
9
Example: Trading Agent
Market
Info
User preferences
Auction
rules
Model of
other market
participants
Strategy
synthesizer
Bidding
strategy
Bids
10
Example: Trading Agent
Beliefs: auction rules, model of market
Desires: user preferences
Intentions: objects it has decided to
buy/sell
Capabilities: place new bids
Observations: market information
11
Reactive Agent
Lookup table maps each observation, or
series of observations, to an action
an = f(on), or
an = f(o1,…, on)
Fast
Inflexible
Intractable for nontrivial domains
12
Rational Agents
Decision theoretic (economic)
Agent makes optimal decisions given its
beliefs, goals, and intentions.
Logical
Agent makes decisions that are consistent
with its beliefs, goals, and intention.
13
Boundedly Rational Agents
Agent makes optimal decisions given its
beliefs, goals, intentions, and the limits of
its computational abilities.
14
Learning Agent
An agent that updates its beliefs based on
its observations
What can we learn?
Model of the world
New capabilities
Effects of our actions
15
Learning Agent
Learning task:
Learn w, where sn+1 = w(sn, an)
Types of learning
Supervised
Reinforcement
Unsupervised
16
Autonomy
Agent autonomy, with respect to
User = execution autonomy
Other agents = social autonomy
Designer autonomy, with respect to
Communication = interface autonomy
Architecture = design autonomy
Utility function = preference autonomy
17
Belief Representation
Knowledge level
The Wolfline runs from HC to CC
Logical level (declarative)
Connects(Wolfline,HC,CC)
Implementation level (procedural)
public class Bus{
public string start = “HC”;
public string end = “CC”;
}
18
Benefits of Declaritivism
Modularity
Semantics
Inspectability
Learnability
Programmability
19
Other Properties of Agents
Lifespan: transient to long-lived
Modeling: of itself, the world, and other
agents
Mobility: stationary or mobile
Memory: non to perfect recall
20
Agent Societies
Software infrastructure
Social organization
21
Software Infrastructure
Communication Channels
Common Ontologies
Service agents (i.e. directory agent)
22
Communication
ACL = agent communication language
Example: KQML
Content messages: tell, query, reply,
etc.
Flow control: next, etc.
Generally do not prescribe semantics
23
Ontologies
Define the semantics of communication
Notoriously difficult
Employee = everyone on payroll
Employee = everyone receiving benefits
24
Social Organization
Homogeneous or heterogeneous
Self-interest v.s. cooperative
Social control structure
System evaluation
25
Emergent Behavior
Agents act
With limited, local knowledge
In self interest
System behaves
In globally desirable manner
Without central control
Adam Smith’s “invisible hand”
26
E-commerce example
Trading agents, again
Heterogeneous
Self-interested
Mediated
Game-theoretic evaluations
27
Legacy System Example
User
Agent
Agent
DB1
Agent
DB2
Agent
DB3
28
Personal Assistants
Agents that support a user’s task
Example (weak)
Dialogue based
Anthropoid
Cooperative ?
But has no goals of its own
29
Personal Assistants
Example: smart calendar/datebook that
could
Negotiate appointments for me
Actively keep track of my contacts by
searching the web
Learn priorities for my mail
30
Conclusion
Agenthood is a convenient description
Agents are described by beliefs, desires, and
intentions
Agents select actions based on observations
Cooperating agents are a form of distributed
computation
Self-interested agents can generate desirable
emergent behavior
31