You are on page 1of 46

Software Agents

ECEG-7521

Multi-Agent Interaction and


Cooperation

Surafel Lemma Abebe (Ph. D.)


Outline
• Interaction
– Coordination
– Cooperation
• Negotiation
• Matchmaking and Brokering

Surafel Lemma Abebe (Ph. D.) 2


Interaction
“The world functions through interacting agents.
Each person pursues his/her own goals through
encounters with other people or machines.”
Rules of Encouter

• Interaction protocols
– Govern the exchange of a series of messages among
agents – a conversation
• Case 1: Agents have conflicting goals
• Case 2: Agents have similar goals

Surafel Lemma Abebe (Ph. D.) 3


Interaction…
• Interaction protocols…
– Agents with conflicting goals
• Self-interested
• Objective of the protocol
– Maximize payoff (utility functions) of agents

– Agents with similar goals


• Agents have common goals or common problems (distributed
problem solving)
• Objective of the protocol
– Maintain globally coherent performance without violating autonomous
behavior of agents
• Main aspects include how to
– Determine shared goals
– Determine common tasks
– Avoid unnecessary conflicts
– Pool knowledge and evidence
Surafel Lemma Abebe (Ph. D.) 4
Interaction…
• Coordination protocols
– Environment may have limited resources
• Agents must coordinate their activities with each other
– Done between multiple agents to satisfy individual or
group goals
– Why action coordination is needed?
• Maintain dependencies between actions
• Meet global constraints
• When no one agent has sufficient competence, resources, or information to
achieve system goals
– Example
• Supplying timely information to other agents, ensuring the actions of agents
are synchronized and avoid redundant problem solving

Surafel Lemma Abebe (Ph. D.) 5


Interactions…
• Coordination protocols…
– Distributed AI (DAI)
• To produce coordinated systems, most DAI researches
focus on distributing data and control
– Distributed control means agents have a degree of autonomy in
generating new actions and in deciding which goals to pursue
next

• Disadvantage of distributing data and control


– Knowledge is distributed throughout the system, so each agent
has only a partial and imprecise perspective
– Increased degree of uncertainty about each agent’s actions
– Difficult to attain coherent global behavior

Surafel Lemma Abebe (Ph. D.) 6


Interactions…
• Coordination protocols…
– Goal graphs
• Actions of agents in solving goals can be expressed as search through a
classical AND/OR goal graph
• Includes a representation of the dependencies between the goals and
resources needed to solve the primitive goals (leaf nodes of the graph)
• Indirect dependencies can exist between goals through shared resources
• Allows the activities requiring coordination to be clearly identified

• Activities to formulate multi-agent systems


1. Defining Goal Graph, including identification and classification of
dependencies
2. Assigning particular regions of the graph to appropriate agents
3. Controlling decisions about which areas of the graph to explore
4. Traversing the graph
5. Ensuring that successful traversal is reported

Surafel Lemma Abebe (Ph. D.) 7


Interactions…
• Coordination protocols…
– Distributed goal search formalism is used to characterize
both global and local problems
– Key agent structures
• Commitment
– Viewed as pledges to undertake a specified course of action
– Provide a degree of predictability
– As situation changes agents must evaluate whether existing
commitments are still valid
– Agents commitment should be both internal and belief consistent
• Convention
– Provide a means to manage commitments in changing circumstances
– Do not specify how the agent should behave towards others, if
commitments are altered or modified
– Provide the necessary degree of mutual support
Surafel Lemma Abebe (Ph. D.) 8
Interaction…
• Coordination protocols…
– Agent structures…
• Convention: Social convention
– For goals that are dependent, its is essential that the relevant agents be
informed of any substantial change that affect them
– Communication resources are limited, the following social convention might
be appropriate

Limited Bandwidth Social Convention


INVOKE WHEN
Local commitment dropped
Local commitment satisfied
ACTIONS
Rule1: IF Local commitment satisfied
THEN inform all related commitments
Rule2: IF Local commitments dropped because unattainable or motivation not present
THEN inform all strongly related commitments
Rule3: IF Local commitments dropped because unattainable or motivation not present
AND communication resources not overburdened
THEN inform all weakly related commitments
Surafel Lemma Abebe (Ph. D.) 9
Interaction…
• Coordination protocols…
– Agent structures…
• Convention: Joint commitment
– Agents decide to pursue a joint action and commit themselves
to a common goal
– Minimum information that a team of cooperating agents should
share
» Status of their commitment to the shared objective
» Status of their commitment to the given team framework
– A change in commitment by one participant could jeopardize
the team’s effort
– Has 3 basic assumptions that represent the minimum
requirement for joint commitment

Surafel Lemma Abebe (Ph. D.) 10


Interaction…
• Coordination protocols…
– Agent structures…
• Convention: Joint commitment…

Basic Joint Action Convention


INVOKE WHEN
Status of commitment to joint action changes
Status of commitment to attaining joint action in present team context changes
Status of joint commitment of a team member changes
ACTIONS
Rule1: IF status of commitment to joint action changes
OR
IF status of commitment to present team context changes
THEN inform all other team members of these changes
Rule 2: IF status of joint commitment of a team member changes
THEN determine whether joint commitment still viable
Surafel Lemma Abebe (Ph. D.) 11
Interaction…
• Cooperation
– Topology
• Independent multi-agent
systems have their own
goals
– Discrete
» Non-correlated goals
– Emergent cooperation
» Agents independent of
each other follow their
goals
» But give the
appearance of a
cooperation

Surafel Lemma Abebe (Ph. D.) 12


Interaction…
• Cooperation…
– Topology…
• Cooperative systems
– Have explicit cooperative mechanisms
– Agents are constructed in a way that they can cooperate with
other agents to achieve their goals
– Agents make intentional use of the capability
– Non-communicative cooperation
» Indirectly using the environment
» Agent observes the environment and responds
– Communicative cooperation
» Agents use communication protocols and procedures to
cooperate with other agents

Surafel Lemma Abebe (Ph. D.) 13


Interaction…
• Cooperation…
– Topology…
• Cooperative systems…
– Communicative cooperation…
» Deliberative systems
• All agents have a common planning and agreement
of the methodology
» Negotiation
• Agents have a competitive component in addition to
the mechanisms of deliberative systems
• Resolves conflicts and assign tasks
• E.g., Contract-Net systems

Surafel Lemma Abebe (Ph. D.) 14


Interaction…
• Cooperation…
– Basic strategy shared by many of the protocols for
cooperation
• Decompose task
• Distribute task
=> Divide and conquer
– Decomposition
• Reduce the complexity of a task
• Smaller sub-tasks require less capable agents and fewer
resources

Surafel Lemma Abebe (Ph. D.) 15


Interaction…
• Cooperation…
– Decomposition…
• Task decomposition could be
– Done by the system designer
– Done by agents using hierarchical planning
– Inherent in the representation of the problem
– Done spatially based on layout of information sources or
decision points
– Done functionally according to the expertise or available
agents

Surafel Lemma Abebe (Ph. D.) 16


Interaction…
• Cooperation…
– Criteria for distributing tasks
• Avoid overloading critical resources
• Assign tasks to agents with matching capabilities
• Make an agent with wide view assign tasks to other agents
• Assign overlapping responsibilities to agents to achieve
coherence
• Assign highly interdependent tasks to agents in spatial or
semantic proximity
– Minimizes communication and synchronization costs
• Reassign tasks if necessary for completing urgent tasks

Surafel Lemma Abebe (Ph. D.) 17


Interaction…
• Cooperation…
– Mechanisms for distributing tasks
• Market mechanisms
– Tasks are matched to agents by generalized agreement or mutual
selection (analogous to pricing commodities)
• Contract net
– Announce, bid, and award cycles
• Multi-agent planning
– Planning agents have the responsibility for task assignment
• Organizational structure
– Agents have fixed responsibilities for particular tasks

Surafel Lemma Abebe (Ph. D.) 18


Interaction…
• Cooperation…
– Partial global planning (PGP)
• Every agent of a multi-agent system is given the capability
during its work to gather information on the current state
and achieved goals of other agents
• Agent uses the collected knowledge to optimize its task

• The knowledge is called partial global plan


– Reflects the partial knowledge of the plan determined by an
agent to solve a global problem

• Condition
– Several distributed agents work on the solution of an overall
problem

Surafel Lemma Abebe (Ph. D.) 19


Interaction…
• Cooperation…
– Partial global planning (PGP)…
• Example: Agents working on two sub problems A and B
– Every agent passes information on its current state to its
opposite partner
– The information is used for its own work
– E.g., for transferring a sub-sub task

Surafel Lemma Abebe (Ph. D.) 20


Interaction…
• Cooperation…
– Partial global planning (PGP)…
• The process is subdivided into the following steps
1. Creation of the local plans for every agent
2. Communication and exchange of plans between the agents
3. Creation of the PGPs
4. Modification and optimization of the PGPs

Surafel Lemma Abebe (Ph. D.) 21


Interaction…
• Cooperation…
– Partial global planning (PGP)…
1. Creation of the local plans for every agent
– The plan is to be used for the solution of the task assigned
– Agent could dynamically adapt its original plan, if newly won
knowledge makes it necessary
– Local plan has two levels of detail
» General structure
• Contains the most important steps to solve the problem
• Also reflects the long-term problem-solving planning
» Detailed structure
• Contains the detailed information for every specific sub-
problem

Surafel Lemma Abebe (Ph. D.) 22


Interaction…
• Cooperation…
– Partial global planning (PGP)…
2. Communication and exchange of plans between the
agents
– Every agent must possess a certain amount of organization-
specific knowledge
» Helps to decide what role the other agents play in the
problem solving process
» Helps to identify for which agent what information could
be of interest
» Agent must decide which parts of its information it passes
to which agents

Surafel Lemma Abebe (Ph. D.) 23


Interaction…
• Cooperation…
– Partial global planning (PGP)…
3. Creation of the PGPs
– Combine received information on the plans to form PGP
– Checks if the new information contains dependencies to its own internal
plans
» Group the associated sub-plans into logical units
– Components of resulting PGP
» Objective
• Contain
• Basic information of a PGP, reason for existence
• Long-term overall objective and the priorities in
comparison with other plans
» Plan activity map
• Covers the tasks of the other agents including their current
state (details of the plans, expected results, and effort
involved)

Surafel Lemma Abebe (Ph. D.) 24


Interaction…
• Cooperation…
– Partial global planning (PGP)…
3. Creation of the PGPs…
– Components of resulting PGP
» Solution construction graph
• Contains information how the individual agents are
to communicate and cooperate with each other
• Details on the size and time of the plans to be sent by
the agents are important here
» Status
• Contains report of all important information for the
PGP
• E.g., references to the plans received by other agents
and an indication of the time of receipt

Surafel Lemma Abebe (Ph. D.) 25


Interaction…
• Cooperation…
– Partial global planning (PGP)…
• Advantage of using PGP
– Highly dynamic behavior of the system
» All plans can be adapted at any time to the new
environmental situation
» Small inconsistencies are acceptable
• Communicates only important modification to plan
• Avoids unnecessary load on the complete system
» System developer is responsible for the development of
tolerant system and the specification of the threshold
value
• Threshold value indicates when a modification can be
considered important
Surafel Lemma Abebe (Ph. D.) 26
Interaction…
• Cooperation…
– Partial global planning (PGP)…
• Advantage of using PGP…
– Efficient and completely avoids redundancy
» Agents working on identical problems would be identified through
their PGP and then cold be reassigned/restructured
» If recourses of an agent are not used fully, it offers an empty plan to
make other agents aware of its free capacity

• Some challenges for the basic approach


– How can heterogeneous agents with different problem solution
strategies be integrated?
– How do you treat dynamic agents whose solution strategies are subject
to repeated changes?
– How can negotiations between agents be implemented on the basis of a
PGP?

Surafel Lemma Abebe (Ph. D.) 27


Interactions…
• Cooperation…
– Black board system
• Characteristics
1. Independence of expertise
» A specialist (knowledge sources, KS) can act
independently of the other
2. Diversity of problem-solving techniques
» Internal representation and inferencing machinery of
each KS is hidden from others
3. Flexible representation of blackboard information
» No restriction as to what can be placed on the
blackboard

Surafel Lemma Abebe (Ph. D.) 28


Interactions…
• Cooperation…
– Black board system
• Characteristics…
4. Common interaction language
» KS’s should be able to correctly interpret information
posted by other KS’S
5. Event based activation
» KS’s are triggered in response to blackboard and external
events
» Each KS informs the blackboard system about the kind of
events in which it is interested

Surafel Lemma Abebe (Ph. D.) 29


Interactions…
• Cooperation…
– Black board system…
• Characteristics…
6. Need for control
» Separate component from the individual KSs
» Responsible for managing the course of problem solving
» Can be viewed as a specialist in directing problem solving
» Triggered KS evaluates quality of its contribution informs
Control Component about the cost estimates benefits
» Control component decides how to proceed for better
problem solving
7. Incremental solution generation
» KSs contribute to the solution as appropriate, sometimes
refining, sometimes contradicting, and sometimes initiating
a new line of reasoning

Surafel Lemma Abebe (Ph. D.) 30


Negotiation
• Aim
– Permit a constructive cooperation from within the group of
independently operating agents that have their own goals

• Definition
– A process by which a joint decision is reached by two or more agents,
each trying to reach an individual goal or objective

• Negotiation protocol provides


– Basic rules for the possible forms of negotiation
– Form of the negotiation process
– The communication basis
• Negotiation strategy depends on the specific implementation of
each agent

Surafel Lemma Abebe (Ph. D.) 31


Negotiation…
• Effects of negotiation
– Different situations can arise when two agents
negotiate with each other
– No synergy effect
• Neither of the agents can achieve an advantage as the
result of the negotiation
• Each agent follows its own specific goal, and there are
no direct dependencies between the two goals
• Neither agent has any improvement to their goals
compared with their opposite party as a result of the
negotiation

Surafel Lemma Abebe (Ph. D.) 33


Negotiation…
• Effects of negotiation…
– Positive synergy effects
• At least one of the two agents achieves its goal faster or
with less effort
• A qualitative improvement of the goal is also possible
• Example
– If the negotiation determines that the goal of one agent is
contained in the goal of the other agent or this goal has been
attained already, this produces a positive synergy effect

Surafel Lemma Abebe (Ph. D.) 34


Negotiation…
• Effects of negotiation…
– Negative synergy effects
• Negotiation is mainly appropriate when there are
conflicting goals
• Could appear when
– The goals of two agents are in direct conflict
» One can achieve the goal
– The goals of two agents cannot be achieved with the available
operations
– The agents resources would be exceeded to achieve a
common goal

Surafel Lemma Abebe (Ph. D.) 35


Negotiation…
• Interactions in negotiations
– Objective of negotiation from an agent’s perspective could be
• Improvement of its own state
• Support for other agents without degrading its own situation
• Requesting a result

– Forms of interaction
• Symmetric cooperation
– Negotiation can produce a result that is better for both agents than each
could achieve by itself
– The work of the other agent has a positive effect on each agent's work
• Symmetric compromise
– Both agents would prefer to achieve their goals independently
– Negotiation is a compromise for both parties
» Causes a degradation of their result
– Both have to negotiate because the existence of the other agent cannot be
ignored

Surafel Lemma Abebe (Ph. D.) 36


Negotiation…
• Interactions in negotiations…
– Forms of interaction…
• Non-symmetric cooperation/compromise
– One of the agents would get a positive result (cooperation) while the
other agent must make a compromise
• Conflict
– Agents cannot agree on a rationally acceptable solution
» Their goals conflict with each other
– Negotiation must be ended without a result being achieved
– Negotiation leads to a result that is unsatisfactory to one of the two
parties
– Which agent leaves the negotiation as winner?
» Randomly select the agent
» Select the agent based on its importance (cost, resource utilization,
…)

Surafel Lemma Abebe (Ph. D.) 37


Negotiation…
• Techniques for negotiation
– Two types
• Environment centered
• Agent centered

– Environment-centered
• Focus
– How can the rule of the environment be designed so that the agent
in it, regardless of their origin, capabilities, or intentions will interact
productively and fairly?
• Attributes of negotiation mechanism
– Efficiency
» Agents should not waste resources in coming to an agreement
– Stability
» No agent should have an incentive to deviate from agreed-
upon strategies
Surafel Lemma Abebe (Ph. D.) 38
Negotiation…
• Techniques for negotiation
– Environment-centered…
• Attributes of negotiation mechanism…
– Simplicity
» Negotiation mechanism should impose low computational
and bandwidth demands on the agents
– Distribution
» Negotiation mechanism should not require a central
decision maker
– Symmetry
» Negotiation mechanism should not be biased against any
agent for arbitrary or inappropriate reasons

Surafel Lemma Abebe (Ph. D.) 39


Negotiation…
• Techniques for negotiation…
– Agent-centered
• Focus
– Given an environment in which my agent must operate, what is
the best strategy for it to follow?
• Such negotiation strategies are developed for specific
problems
• Two general approaches
• 1st approach
– Uses speech-act classifiers with a possible world semantics to
formalize negotiation protocols and their components
» Clarifies the conditions of satisfaction for different kinds of
messages

Surafel Lemma Abebe (Ph. D.) 40


Negotiation…
• Techniques for negotiation…
– Agent-centered…
• 2nd approach
– Unified negotiation protocol
– Assumes
» Agents are economically rational
» Set of agents is small
» Agents have a common language
» Agents have common problem abstraction
» Agents must reach a common solution
– Agents that follow this protocol create a deal, i.e., a joint plan
between the agents that would satisfy all of their goals

Surafel Lemma Abebe (Ph. D.) 41


Matchmaking and brokering
• An agent may not have the necessary knowledge
or resources to perform a task
• Challenge
– Efficient search for suitable agents to solve a specific
problem
• Solution
– Delegate a task to another agent
• If the agent knows a suitable partner, it doesn’t require
external assistance
• Otherwise, it is dependent on external support
– Matchmakers or brokers

Surafel Lemma Abebe (Ph. D.) 42


Matchmaking and brokering…
• Matchmaker
– Is an agent with special capabilities that has the task to match
information-searching and information-providing agents
– Possesses a database with information on the information-
providing agents
• Every agent reports its service with a matchmaker
• Server/provider
– Service providing agent
– Searches its database for suitable servers and passes their
names to the requester

Surafel Lemma Abebe (Ph. D.) 43


Matchmaking and brokering…
• Matchmaker…
– Inquiry by the matchmaker doesn’t represent a binding
commitment for a requester
– When server reports the service availability, it also makes the
binding commitment to perform the corresponding task
– Requirements
• Name or address of the matchmaker must be known to all agents
• Every agent must have the authorization to contact the matchmaker
– To make an inquiry
– To provide its services
• Matchmaker database should store at least
– Name of agent offering the service
– Type of offered service (problem or sub-problem to be solved)
– Costs or resources required to solve the problem
– Reliability of the agent

Surafel Lemma Abebe (Ph. D.) 44


Matchmaking and brokering…
• Broker
– If an agent makes an inquiry to the broker, it expects the
name of the agent and solution
– Task of broker
• Find a provider to solve the problem
• Negotiate with the provider
• Pass the task to the provider
• Forward the result back to the inquiring agent

Surafel Lemma Abebe (Ph. D.) 45


Matchmaking and brokering…
• Broker…
– A combination of both concepts is possible
• Example
– Use a matchmaker to find a broker
– Requester enters a fixed commitment with the inquiry
to a broker
• Commissions the broker to delegate the requested task
• Guarantees that it will accept all costs or resource
consumption that occur
– From requester point of view, no difference between
the broker and the server

Surafel Lemma Abebe (Ph. D.) 46


Matchmaking and brokering…
• Broker…
– Advantage
• Better resource distribution
– Broker has total knowledge of other agents
– Agents report to the broker
» Helps the broker to know which agents are overloaded or
underloaded and guarantee an efficient load
– Broker could recruit another agent (broker)
• Result is to be sent directly to the requester and not
the recruiting broker
– Reduces the communication and resource loads

Surafel Lemma Abebe (Ph. D.) 47

You might also like