1. From Multiagent Systems to Multiagent Societies
Michael Berger
Based on:
1) “Multiagent Systems and Societies of Agents” / Michael N. Huhns and Larry M. Stephens
In Multiagent Systems: A Modern Approach to Distributed Artificial Intelligence (Chapter 2) / Gerhard W. Weiss
2) “Commitments and Conventions: The Foundation of Coordination in Multi-Agent Systems” / Nick R. Jennings

2. Overview Agent and Environment Communications Interactions
Commitments and Conventions

3. Part I: Agent and Environment

4. Agent - Definition
An active object with the ability to perceive, reason and act.
Has explicitly represented knowledge and a mechanism for operating on or drawing inferences from its knowledge.
Has the ability to communicate.

5. Environment - Categories
Knowable (Accessible)
Predictable (Deterministic)
Controllable
Historical (non-Episodic)
Telelogical
Real-time (Dynamic) No
Yes
Open

6. Part II: Communications

7. Communications - Overview
Motivation
Meanings
Speech Acts
Message Types and Dialogue Roles
Communication Protocols
KQML
KIF
Ontologies

8. Motivation (I)
Coordination - the extent to which agents avoid extraneous activity.
Reducing resource contention
avoiding livelock / deadlock
maintaining safety conditions
Coherence - how well the system behaves as a unit.
Determining shared goals
Pooling knowledge and evidence ?

9. Motivation (II) Coordination - “not making things worse”.
Coherence - “making things better”.
Communication enables the agents to coordinate their actions and behavior, resulting in systems that are more coherent.

10. Meanings (I) Communication - consists of:
Syntax - how the symbols of communication are structured.
Semantics - what the symbols denote.
Pragmatics - how the symbols are interpreted.
Meaning = Semantics + Pragmatics

11. Meanings (II) Dimensions of meaning: Descriptive vs. Prescriptive
Speaker’s vs. Hearer’s vs. Society’s Perspective
Semantics vs. Pragmatics
Contextuality
Identity
Cardinality

12. Speech Acts
Speech act theory used as basis for analyzing human communication.
Theory views human natural language as actions.
Speech acts have three aspects:
Locution - the physical utterance by the speaker.
Illocution - the intended meaning of the utterance by the speaker.
Perlocution - the action that results from the locution.
“Performative” - Speech acts that have the property that “saying it makes it so” (e.g. promise, report, tell, request, demand).

13. Message Types and Dialogue Roles
Two basic message types:
Assertion
Query
Three dialogue roles:
Master (active)
Sends queries (questions), receives assertions (answers), sends assertions (fact determinations).
Slave (passive)
Receives queries (questions), sends assertions (answers), receives assertions (fact determinations).
Peer
Master + Slave

14. Communication Protocols
Communication can be:
binary (single sender, single receiver)
n-ary (single sender, many receivers)
Messages sent using communication protocols are specified by a data structure, that contains the following fields:
Sender
Receiver
Encoding / Decoding functions
Language of message
Message content

15. Communicating Agents (I)
a is broken.

16. KQML KQML - Knowledge Query and Manipulation Language.
Basic KQML performative defined by a structure that contains the following fields:
Sender
Receiver
Language
Ontology
Content
More advanced performatives.
Language used as wrapper for other languages - Domain independent!
Forwarding and nesting possible.

17. Communicating Agents (II)
Je ne comprends pas
Sender: Cowboy
Receiver: Shadow
Language: English
Content: a is broken
Languages: English,
Spanish, Basque
Languages: French

18. KIF KIF - Knowledge Interchange Format.
Prefix version of first-order predicate calculus.
Example: or ((and (> ?a 6) (> b 5))) (< c 7)
Possible to encode knowledge about knowledge (second-order) and to describe procedures.

19. Communicating Agents (III)
Sender: Cowboy
Receiver: Shadow
Language: KIF
Ontology: Computers
Content: broken(a)
bad(message)
Languages: English,
Spanish, Basque, KIF
Languages: French, KIF
Ontologies: Fashion,
Politics, Weather
Ontologies: Computers,
Politics, Sports

20. Ontologies
Ontology - specification of objects, concepts and relationships in an area of interest (domain).
Concepts represented in first-order logic as unary predicates. Relationships represented by n-ary predicates.
Note: predicates refer to classes of objects, not instances of objects.
except “instanceof”
All agents share the same ontology - i.e. all agents use and understand the same “vocabulary”!

21. Communicating Agents (IV)
Sender: Cowboy
Receiver: Shadow
Language: KIF
Ontology: Computers
Content: broken(a)
need_fixing(a)
Computer Ontology:
instanceof(a, disk)
instanceof(X, disk) AND broken(X) ==> need_fixing(X)
Languages: English,
Spanish, Basque, KIF
Languages: French, KIF
Ontologies: Fashion,
Politics, Weather, Computers
Ontologies: Computers,
Politics, Sports

22. Part III: Interactions

23. Interactions - Overview
Motivation
Negotiation
Market Mechanisms
Contract Net
Truth Maintenance Systems
Blackboard Systems

24. Motivation
Communication is a necessary condition for coordination and coherence, but not a sufficient one.
It would help if agents could:
Determine shared goals
Avoid unnecessary conflicts
Pool knowledge and evidence

25. Negotiation
Negotiation - a process by which a joint decision is reached by two or more agents, each trying to reach an individual goal.
Main steps:
One of the agents communicates its initial position.
While no agreement is reached, each agent makes a proposal in its turn. These may include:
Concessions.
New alternatives.
Ends with agreement or disagreement.
Mechanisms for negotiation may be:
Environment-centered
Agent-centered

26. Negotiation Mechanisms: Environment-Centered
Environment designer.
“How can the rules of the environment be designed so that the agents will interact productively and fairly?”
A negotiation mechanism would ideally have the following attributes:
Efficiency
Stability
Simplicity
Distribution
Symmetry

27. Negotiation Mechanisms: Agent-Centered
Agent designer.
“Given an environment, what is the best strategy for my agent to follow?”
Large part of the negotiation mechanisms assume that agents are economically rational.
For example, a negotiation protocol that contains the following terms:
Deal
Utility
Negotiation set

28. Market Mechanisms (I)
Everything of interest to the agents described in terms of prices.
Two types of agents:
Consumers
Producers
Markets of goods are interconnected.

29. Market Mechanisms (II)
Big market will usually reach a competitive equilibrium:
Consumers bid to maximize utility, subject to their budget constraints.
Producers bid to maximize profits, subject to their technological capability.
Net demand is zero for all goods.

30. Contract Net (I) Interaction protocol for cooperative problem solving.
Modeled on the contracting mechanism used by businesses.
For any assignment, agents are divided ad-hoc into managers and contractors.

31. Contract Net (II) Managers: Contractors:
Announce a task that needs to be performed.
Receive and evaluate bids from potential contractors.
Award a contract to a suitable contractor.
Receive and synthesize results.
Contractors:
Receive task announcements.
Evaluate their own capability to respond.
Respond (decline / bid).
Perform the task if bid is accepted by manager.
Report task’s results. 1 5 6 9 2 3 4 7 8

32. Truth Maintenance System (I)
Truth Maintenance System (TMS) - ensures the integrity of an agent’s knowledge, and keeps the knowledge base:
Stable
Each datum that has a valid justification is believed.
Each datum that lacks a valid justification and which is not in initial belief set is disbelieved.
Well-founded
Permits no set of its beliefs to be mutually dependent.
Logically consistent
No datum is both believed and disbelieved.
Every datum is either believed or disbelieved.
No data and its negation are both believed or disbelieved.

33. TMS Graph _ + _ _ + _ U (OUT) T (EXTERNAL) P (IN) T (INTERNAL) Q (OUT)
Agent 2 +
P (IN)
T (INTERNAL) _ _
Q (OUT)
R (IN)
S (OUT) + _
Agent 1

34. Truth Maintenance System (II)
Every datum is labeled either:
IN (in initial belief set).
INTERNAL (“IN” because of local justification).
EXTERNAL (“IN” because another agent asserts it).
OUT (disbelieved).
When justification is added or removed, the TMS is invoked:
Some data unlabeled, including the newly justified datum and its consequences in all agents.
New Labeling introduced for all unlabeled data.
If any affected agent fails to label, backtrack occurs.
Principal of TMS changes: Affect as few agents as possible and as few beliefs as possible.

35. TMS - Example (I) _ + _ _ + _ U (OUT) T (EXTERNAL) P (IN) T (INTERNAL)
Agent 2 +
P (IN)
T (INTERNAL) _ _
Q (OUT)
R (IN)
S (OUT) + _
Agent 1

36. TMS - Example (II) _ + + _ _ + _ U (OUT) T (EXTERNAL) P (IN)
Agent 2 +
P (IN)
T (INTERNAL) + _ _
Q (OUT)
R (IN)
S (OUT) + _
Agent 1

37. TMS - Example (III) _ + + _ _ + _ U (OUT) T (EXTERNAL) P T (INTERNAL)
Agent 2 + P
T (INTERNAL) + _ _ Q
R (IN)
S (OUT) + _
Agent 1

38. TMS - Example (IV) _ + + _ _ + _ U T P T Q R (IN) S (OUT) Agent 2

39. TMS - Example (V) U T _
Agent 2 + P T + _ _ Q R S + _
Agent 1

40. TMS - Example (VI) _ + + _ _ + _ U (IN) T (OUT) P (OUT) T (OUT)
Agent 2 +
P (OUT)
T (OUT) + _ _
Q (OUT)
R (OUT)
S (IN) + _
Agent 1

41. Blackboard Systems (I)
Akin to the following metaphor:
A group of specialists working together on solving a problem.
A common blackboard allows every specialist to report (“write down”) his sub-task results.
Every specialist may be assisted in his work by information reported on the blackboard.
Every specialist is called a “knowledge source” (KS).

42. Blackboard Systems (II)
Characteristics of blackboard systems:
Independence of expertise.
Diversity in problem-solving techniques.
Flexible representation of blackboard information.
Common interaction language.
Event-based activation.
Need for control.
Incremental solution generation.

43. Part IV: Commitments and Conventions

44. Distributed Goal Search Model
Goals solution expressed as AND/OR graph (which is directed and a-cyclic).
High-level goals are root nodes.
Primitive goals are leaf nodes.
Graph also contains resources needed for solving primitive goals.
Dependencies may exist between different goals or between a goal and its resource.
Strong vs. weak
Uni-directional vs. bi-directional
Note that dependencies from resources to goals may be solved by adding more instances of the resource.

45. Distributed Goal Search Graph
Agent1
Agent2 G1 G2
G11
G12
……….…
G1k
G1,2m
G2p
………………
G2t
Strong dependencies
G11,1
G11,2
G1m,1
G2m,2
G2p,1
G2p,2
Weak dependendcies
G1m,1,1
G1m,1,2
G2p,1,1
G2p,1,2
G2p,2,2
d11
d1j
d2j+1
d2z
……………………………………
……………………………
Resources
Goals

46. Interactions among Agents for Distributed Goal Search
Defining the goal graph.
Assigning particular regions of the graphs to different agents.
Controlling decisions about which areas of the graph to explore.
Traversing the graph.
Ensuring that successful traversal of the graph is reported.

47. Commitment - Definition
A pledge from one agent to another agent (or itself) to undertake a specified course of action.

48. Commitments
Practical reasoning agents employ intentions for choosing a course of action - a kind of “self-commitment”.
In computational problems, different agents commit themselves to solving different sub-goals of a larger goal.
Agents may inform other agents of the sub-goals to which they are self-committed. In stronger terms, they may commit to other agents about solving these sub-goals.

49. Motivation for Conventions (I)
Agents do not have complete knowledge of the goals and intentions of other agents.
Infeasible to have all agents re-contemplate about the goals of other agents in every step:
Limited computation power
Limited communication bandwidth
Infeasible to have one agent or database keep all information about all agents:
Bottleneck
Single point of failure

50. Motivation for Conventions (II)
If circumstances changed, an agent might be working sub-optimally until he asks about it.
Another agent solves a goal
Another agent commits itself to a goal
Another agent drops his commitment to a goal
Another agent discovers that a goal is no longer attainable
We still would like to keep a distributed system of agents...

52. Convention - Definition
A pre-determined description, common to all agents in the system, of the course of action to be taken by an agent, given a specific circumstance or occurrence.

53. Minimum Convention for Joint Commitments
Formalism by Cohen and Levesque.
BASIC SOCIAL CONVENTION
REASONS FOR ACTION:
STATUS OF COMMITMENT TO SHARED GOAL CHANGES
STATUS OF COMMITMENT TO REACHING SHARED GOAL IN PRESENT TEAM CONTEXT CHANGES
STATUS OF COMMITMENT OF A TEAM MEMBER TO SHARED GOAL CHANGES
ACTIONS:
R1: IF STATUS OR COMMITMENT TO SHARED GOAL CHANGES OR
STATUS OF COMMITMENT IN PRESENT TEAM CONTEXT CHANGES
THEN INFORM ALL OTHER TEAM MEMBERS OF CHANGE
R2: IF STATUS OF COMMITMENT OF A TEAM MEMBER TO SHARED GOAL CHANGES
THEN DETERMINE WHETHER JOINT COMMITMENT STILL VIABLE

54. Convention for Limited-Bandwidth Environments
LIMITED-BANDWIDTH CONVENTION
REASONS FOR ACTION:
COMMITMENT SATISFIED
COMMITMENT UNATTAINABLE
MOTIVATION FOR COMMITMENT NO LONGER PRESENT
ACTIONS:
R1: IF COMMITMENT SATISFIED OR
COMMITMENT UNATTAINABLE OR
THEN DROP COMMITMENT
R2: IF COMMITMENT SATISFIED
THEN INFORM ALL AGENTS WORKING ON RELATED GOALS
R3: IF COMMITMENT DROPPED BECAUSE UNATTAINABLE OR
MOTIVATION NOT PRESENT
THEN INFORM ALL AGENTS WORKING ON STRONGLY RELATED GOALS
R4: IF COMMITMENT DROPPED BECAUSE UNATTAINABLE OR
MOTIVATION NOT PRESENT AND COMMUNICATION RESOURCES NOT OVERBURDENED
THEN INFORM ALL AGENTS WORKING ON WEAKLY RELATED GOALS

55. Convention in Nearly Open Environments (I)
JOINT RESPONSIBILITY SOCIAL CONVENTION
INHERIT: BASIC SOCIAL CONVENTION
REASONS FOR ACTION:
SHARED GOAL IS MET
SHARED GOAL WILL NEVER BE MET
MOTIVATION FOR SHARED GOAL IS NO LONGER PRESENT
AGREED PLAN WILL NOT ACHIEVE DESIRED RESULTS
AGREED PLAN CANNOT BE EXECUTED
AGREED PLAN HAS NOT BEEN EXECUTED PROPERLY
ACTIONS:
R1: IF SHARED GOAL IS MET OR
SHARED GOAL WILL NEVER BE MET OR
THEN DROP COMMITMENT TO SHARED GOAL & TO AGREED PLAN
R2: IF AGREED PLAN WILL NOT ACHIEVE DESIRED RESULTS OR
AGREED PLAN CANNOT BE EXECUTED OR
THEN DROP COMMITMENT TO AGREED PLAN
R3: IF DROP JOINT COMMITMENT TO AGREED PLAN AND
CAN RE-PLAN USING SAME AGENTS
THEN DEVELOP AND COMMIT TO NEW PLAN

56. Convention in Nearly Open Environments (II)
R4: IF DROPPED COMMITMENT TO AGREED PLAN AND
CANNOT RE-PLAN USING SAME AGENTS AND
CAN DEVELOP NEW PLAN USING DIFFERENT TEAM
THEN DROP COMMITMENT TO EXISTING TEAM & COMMIT TO NEW TEAM
R5: IF CANNOT DEVELOP NEW COMMON PLAN
THEN DROP COMMITMENT TO SHARED GOAL & TO AGREED PLAN

57. Possible Trends in Conventions
The harsher the environment, the more rules are needed to determine the agent’s action.
The harsher the environment, the more frequent are situations in which the agent stops and reconsiders objectives.
Similar to the spectrum between bold agents and cautious agents (Kinny and Georgeff).

58. Example for Benefits of Conventions
Agent1
Agent2 G1 G2
G11
G12
……….…
G1k
G1,2m
G2p
………………
G2t
G11,1
G11,2
G1m,1
G2m,2
G2p,1
G2p,2
G1m,1,1
G1m,1,2
G2p,1,1
G2p,1,2
G2p,2,2
d11
d1j
d2j+1
d2z
……………………………………
……………………………
Resources
Goals
Strong dependencies
Weak dependendcies

59. Agents without Honor
G1,21
Agent 1: G11,1 Agent 2: G21,2 Agent 1 reneges
G11,1
G21,2
G1,21
Agent 1: G11,1 Agent 2: G2p Agent 1 reneges
G11,1
G21,2
G2m
Agent 1: G1k Agent 2: G2m Agent 1 reneges
G1k
G2m,1
G2m
Agent 1: G1k Agent 2: G2m Agent 1 reneges
G1k
G2m,1
G2m,2
G2m
G1k
Agent 1: G1k,2 Agent 2: G2m Agent 2 reneges
G1k,1
G1k,2
G2m,1

60. Benefits of Conventions
Provides a degree of predictability to counteract the uncertainty caused by the distribution of control.
Mitigates the effect of commitments reneged.
Flexible - can sometimes be made at different levels and thus have varied time horizons.
The lower the level, the higher the accuracy of information and the larger are the required computation and communication bandwidth
Lower levels don’t always provide a significant contribution
Lower levels might cause more constraints

61. Conventions vs. Conventions
Humans also have conventions.
Not obligatory.
Others don’t always expect adherence to them.
Agent conventions are actually rules rather than conventions.

62. The Dawn of Society
In a System of agents/humans acting without conventions, they cannot expect anything from their peers.
Pre-determined rules/conventions act as common denominator for all units.
Conventions that are adhered to, allow the system to act more coherently without extra effort from particular units.
The whole is larger than the sum of its parts.
Thus a system turns into a society.
Human societies always have unwritten rules.
Agent conventions also called “social rules”.

63. Thank You