Transcript Document

Towards Decentralization
by
Matti Saastamoinen
12.02.2003
Contents
 Introduction
 Multiagent systems
 Distriputed problem solving and planning
 Distributed Rational Decision Making
 Conclusions
12.02.2003
Introduction
 Book: Multiagent Systems, A Modern Approach to
Distributed Artificial Intelligence
 edited by Gerhard Weiss, TUM
 written year 1999
 many (22) authorities in the book
 introductory text and a textbook that covers the
whole range of multiagent systems
 key concepts, methods and algorithms that form the
core of the field
 extensive glossary
12.02.2003
Introduction
 artificial intelligence (AI) is the branch of computer
science concerned with making computers behave like
humans
 the term was first used in 1956 by John McCarthy at
the Massachusetts Institute of Technology
 expert systems in the early 1980s
 the study of multiagent systems began in the field of
distributed artificial intelligence (DAI) about 20 years
ago
 greatest advances have occurred in the field of
games playing (Deep Blue and Deep Junior)
 most common are LISP and Prolog
12.02.2003
Introduction
 artificial intelligence includes:
• games playing: programming computers to play
games such as chess and checkers
• expert systems : programming computers to make
decisions in real-life situations
• natural language : programming computers to
understand natural human languages
• neural networks : Systems that simulate
intelligence by attempting to reproduce the types of
physical connections that occur in animal brains
• robotics : programming computers to see and
hear and react to other sensory stimuli
12.02.2003
Multi-agent systems, motivations
 distributed computations are sometimes easier to
understand and easier to develop, especially when the
problem being solved is itself distributed
 there are also times when a centralized approach is
impossible, because the systems and the data belong
to independent organization that want to keep their
information private and secure for competitive reason
 for the practical reason that the systems are too
large and dynamic for global solutions to be
formulated and implemented, the agents need to
execute autonomously and be developed
independently
12.02.2003
Characteristics of environments
12.02.2003
Agent communications
 agents to communicate single messages
 agents communicate in order to achieve better the
goals of themselves or of the society/system in which
they exits
 communication can enable the agents to coordinate
their actions and behavior, resulting in systems that
are more coherent
 coordination is a property of a system of agents
performing some activity in a shared environment
 cooperation is coordination among no antagonistic
agents
 negotiation is coordination among competitive or
simply self-interested agents
12.02.2003
Agent communications
 a taxonomy of the different ways in which agents
can coordinate their behavior and activities
12.02.2003
Agent communications
 three aspects to the formal study of communication:
• syntax, how the symbols of communication are
structured
• semantics, what the symbols denote
• pragmatics, how the symbols are interpreted
12.02.2003
KQML
 Knowledge Query and Manipulation Language
(KQML) is a protocol and language for exchanging
information and knowledge
 KQML-speaking agents appear to each other as
clients and servers
 KQML is a protocol for communications among both
agents and application programs
12.02.2003
KIF
 agents need descriptions of real-world things
 Knowledge Interchange Format (KIF), a particular
logic language, has been proposed as a standard to
use to describe things with in expert systems,
databases, intelligent agents, etc.
 it is readable by both computer systems and people
 the expression shown below is an example of a
complex sentence in KIF. It asserts that the number
obtained by raising any real number ?x to an even
power ?n is positive:
(=> (and (real-number ?x)
(even-number ?n))
(> (expt ?x ?n) 0))
12.02.2003
Agent interaction protocols
 interaction protocols govern the exchange of a
series of messages among agents – a conversation
 agents can have conflicting goals or are simply selfinterested, the objective of the protocols is to
maximize the payoffs of the agents
 agents may have similar goals or common
problems, the objective of the protocols is to maintain
globally coherent performance of the agents without
violating autonomy
 coordination and cooperation protocols
12.02.2003
Coordination protocols
 agents must coordinate their activities with each
other to further their own interests or satisfy group
goals
 there are dependencies between agents’ actions
 a need to meet global constraints
 no one agent has sufficient competence, resources
or information to achieve system goals
 usually data and control is distributed
 disadvantage of distributing control and data is that
knowledge of the system’s overall state is dispersed
throughout the system
 commitments are viewed as pledges to undertake a
specified course of actions
12.02.2003
Coordination protocols
 conventions provide a means of managing
commitments in changing circumstances
 commitments and conventions are the cornerstones
of coordination
12.02.2003
Cooperative protocols
 a basic strategy is to decompose and then distribute
tasks
 task decomposition might be done spatially, based
on the layout of information sources or decision
points, or functionally, according to the expertise of
available tasks
12.02.2003
Cooperative protocols
 decomposed tasks distribution criteria:
• avoid overloading critical resources
• assign tasks to agents with matching capabilities
• make an agent with a 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. This minimizes
communication and synchronization costs
• reassign tasks if necessary for completing urgent
tasks
12.02.2003
Cooperative protocols
 commonly used tasks distribution mechanisms are:
• 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
responsibility for particular tasks.
12.02.2003
Distributed problem solving and planning
 emphasis is on getting agents to work together well
to solve problems that require collective effort
 coherence: agents need to want to work together
 competence: agents need to know how to work
together well
 distributed planning is tightly intertwined with
distributed problem solving, being both a problem in
itself and a means to solving a problem
12.02.2003
Motivations
 using distributed resources concurrently can allow
a speedup of problem solving thanks to parallelism
 expertise or other problem-solving capabilities can
be inherently distributed.
 beliefs or other data can be can be distributed
 the results of problem solving or planning might
need to be distributed to be acted on by multiple
agents
12.02.2003
Task Sharing
 decomposition: Generate the set of tasks to
potentially be passed others. This could generally
involve decomposing large tasks into subtasks that
could be tackled to different agents
 allocation: Assign subtasks to appropriate agents
 accomplishment: The appropriate agents each
accomplish their subtasks, which could include
further decomposition and subsubtask assigment,
recursively to the point that an agent can
accomplish the task it is handed alone
 result synthesis: When an agent accomplish its
subtask, it passes the result to the appropriate
agent, who knows how to compose results into an
overall solution
12.02.2003
Result Sharing
 can improve group performance in the following
ways:
• Confidence: Independently derived results for the
same task can be used to corroborate each other,
yielding a collective result that has higher
confidence of being correct
• Completeness: Each agent formulates results for
whichever subtasks it can accomplish, and these
results altogether cover a more complete portion of
the overall task
• Precision: To refine its own solution, an agent
needs to know more about the solutions that others
have formulated
12.02.2003
Result Sharing
• Timeliness: Even if an agent could in principle
solve a large task alone, solving subtasks parallel
can yield an overall solution faster
 difficulties in result sharing:
• agents need to know what to do with shared
results
• communicating large volumes of results can be
costly
• as selective as possible about what to exchange
12.02.2003
Distributed Planning
 is something of an ambiguous term, because it is
unclear exactly what is ‘distributed’, the planning or
plans
 can be following kind of distributions:
• centralized planning for distributed plans
• distributed planning for centralized plans
• distributed planning for distributed plans
12.02.2003
Distributed Planning and Execution
 post-planning coordination means that if one agent
hits conflict (can’t reach the goals) there are two ways
to solve this conflict
• each agent formulates not only its expected plan,
but also alternative plans to respond to possible
contingencies that can arise at execution time
• through monitoring and replanning: Each agent
monitors its plan execution, and if there is a
deviation it stops all agent’s progress, and the plancoordination-execution cycle is repeated
 pre-planning coordination means that before an
agent begins planning at all the conflict situations are
found out
12.02.2003
Distributed Rational Decision Making
 automated negotiation systems with self-interested
agents are becoming increasingly important
 each agent is trying to maximize its own good
without concern for the global good
 save labor time of human negotiators
 other savings are possible because computational
agents can be more effective at finding beneficial shortterm contracts
 protocols need to be designed using a noncooperative, strategic perspective
 robust non-manipulable multiagent systems, where
the agents may be constructed by separate designer
and/or may represent different real world parties.
12.02.2003
Evaluation Criteria
 negotiation protocols – i.e. mechanisms – can be
evaluated according to many types of criteria
 the choice of protocol will then depend on what
properties the protocol designer wants the overall
system to have
 Social Welfare
 Pareto Efficiency
 Individual Rationality
 Stability
 Computational Efficiency
 Distribution and Communication Efficiency
12.02.2003
Voting
 plurality protocol
 binary protocol
 Borda protocol
12.02.2003
Auctions
 many practical computer science applications
 web sites exist for buying and selling items using
auction protocols
 deal between two agents: the auctioneer and one
bidder
 auction settings: private, common and correlated
value auctions
 auction protocols:
• English (first-price open-cry) auction
• the first-price sealed-bid auction
• Dutch (descending) auction
• Vickrey (second-price sealed-bid) auction
 all-pay auctions
 lookahead when auctioning items one at a time
12.02.2003
Bargaining
 agents can make a mutually beneficial agreement,
but have a conflict of interest about which agreement
to make
 axiomatic bargaining theory does not use the idea of
a solution concept where the agent’s strategies form
some type of equilibrium. Instead, desirable properties
for a solution, called axioms of the bargaining
solution, are postulated, and then the solution concept
that satisfies these axioms is sought (Nash bargaining
solution).
 strategic bargaining theory the bargaining situation
is modeled as a game, and the solution concept is
based on an analysis of which of the player’s
strategies are in equilibrium
12.02.2003
General Equilibrium Market Mechanisms
 successfully adapted for and used in computational
multiagent systems in many application domains
 provides a distributed method for efficiently
allocating goods and resources
 two types of agents: consumers and produces
 actual production and consumption only occur once
the market has reached a general equilibrium
 most common decentralized algorithm for
equilibrium search is the price tâtonnement process,
which is a steepest descent search method
 use a single centralized mediator
 mediator might become a communication and
computation bottleneck or a potential point of failure
for the whole system
12.02.2003
Contract Nets
 formal model to a negotiation protocol that provably
leads to desirable task allocation among agents
 contracting decisions are based on marginal cost
calculations
 a contract is individually rational (IR) to an agent if
that agent is better off with the contract than without it
 marginal cost is dynamic, it depends on the other
tasks that the contractor already has
 a contractor is willing to allocate the task from its
current task set to the contractor if it has to pay the
contractor less than it save by handling the task itself
 accepting/rejecting decisions based on these
marginal cost calculations
 the task allocation can only improve at each step
12.02.2003
Coalition Formation
 coalition formation is often studied in a more
abstract setting called a characteristic function game
(CFG)
 the value of each coalition is given by a
characteristic function
 Coalition formation in CFGs includes three
activities:
• Coalition structure generation, with three agents, there
are seven possible coalitions: {1}, {2}, {3}, {1,2}, {2,3}, {1,3}, {1,2,3}
and five possible coalition structures: {{1}, {2}, {3}}, {{1}, {2,3}}, {2,
{1,3}}, {{3}, {1,2}}, {{1,2,3}}.
• Solving the optimization problem of each coalition
• Dividing the value of the generated solution
among agents
12.02.2003
Conclusions
 distributed intelligent agents have a significant role
in the future of software engineering!?
 further research is needed to develop the basis and
techniques for societies of computational agents
 distributed planning has a variety of reasonable
well-studied tools and techniques in it repertoire
 challenges is in characterizing these tools and
understanding where and when to apply each
 goal of having heterogeneous plan generation and
plan execution agents work together is likely to remain
elusive
 representations and general-purpose strategies for
distributed problem solving are even more elusive
12.02.2003
Conclusions
 distributed problem solving and planning strategy
has still more ‘art’ to it than we like to see in an
engineering discipline
 real-time search provides an attractive framework,
but there are still unresolved problems
 in the future, systems will increasingly be designed,
built, and operated in a distributed manner
 a large number of systems will be used by multiple
real-world parties (today internet)
 the problem of coordinating these parties and
avoiding manipulation cannot be tackled by
technological or economic methods alone
 the successful solutions are likely to emerge from a
deep understanding and careful hybridization of both
12.02.2003
Conclusions
 centralized systems are failing for two simple
reasons: They can't scale, and they don't reflect the
real world of people
 decentralization is neither automatic nor absolute
 the most decentralized system doesn't always win
 the challenge is to find the equilibrium points--the
optimum group sizes, the viable models and the
appropriate social compromises
12.02.2003
Conclusions
 Modularity + Decentralization -> Changeability
12.02.2003
thank you for your attention!
12.02.2003