MultiAgent%20Architecture%20and%20an%20Example

1
MultiAgent Architecture and an Example
2

• Ana Lilia Laureano-Cruces
• e-mail clc_at_correo.azc.uam.mx
• http//delfosis.uam.mx/ana/AnaLilia.html
• Universidad Autónoma Metropolitana Azcapotzalco
  - MEXICO

3
Distributed Artificial Intelligence

• Distributed resolution of problems
• MultiAgent systems

4
Distributed resolution of problems

• Cooperating modules or nodes
• The knowledge about the problem and the
  development of the solution is distributed

5
MultiAgent Systems

• Coordinated intelligent behaviour between a
  coordinated collection of autonomos agents
• Knowledge
• Goals
• Skills
• Planning
• Reasoning about the coordination between agents

6
Contents

• Basic ideas
• Introduction (Control Theory and Cognitive
  Psychology)
• MultiAgent Systems
• An expert decision application
• Conclusions

7
Basic Ideas

• The intelligence of the majority of traditional
  problem solving algorithms is incoporated by the
  designer.
• As a result, they are predictable and do not
  allow for unexpected results.
• This type of systems are repetitive, and always
  yield the same output for a given set of input
  data.
• Modifying these codes is normally a very
  complicated task.

8
Basic Ideas

• The resolution methods based on the association
  of agents are conceived to exhibit emergent
  behavior rather than a predicatble one.
• It is possible to create new agents to take care
  of situations that are not taken into
  consideration during the original design, without
  the need of modifying existing agents.
• The basic idea is to conceive the solution as a
  set of restrictions to be satisfied rather than
  as the result of a search process.

9
Basic Ideas

• By creating a society of agents, it is possible
  that each one of them is in charge of a subset of
  restrictions.
• In this manner, the global problem is solved
  through a series of negotiations or intervention
  hierarchy between agents, rather than through
  searching.
• Each agent could represent different interest
  conflicts, which should be followed carefully.
• If at the end of the iteration an adequate
  solution is not reached, a restriction has not
  been taken into account, and an agent that
  considers it should be introduced.

10
The nature of AI problems

• There are two classes of AI problems.
• Classic problems (related with optimization).
• Everyday problems of human beings.
• The central idea is to find a solution that,
  without being optimum, satisfies our requirements.

11
When we think in MultiAgent Systems to solve the
problem we most take into account some ideas ...

• In spite of its complexity, any problem can be
  decomposed in tractable parts.
• The relationship between its parts is weak, that
  is, an increasing complexity does not affect the
  interaction between them.
• The specifications of the problem and the control
  is distributed among all the agents.

12
When we think in MultiAgent Systems to solve the
problem we most take into account some ideas ...

• An individual agent is not interested in the
  global problem it is solving.
• The result of the interaction of agents provides
  the solution that is being searched.
• This perspective is that of distributed AI.

13
When we think in MultiAgent Systems to solve the
problem we most take into account some ideas ...

• What is the difference between the classical and
  agent strategies?
• S (p1,p2,...pk).
• S p1 x p2 x ... x pk
• S p1 x p2

14
When we think in MultiAgent Systems to solve the
problem we most take into account some ideas ...

• The problem is distributed.
• Each agent represents a relevant entity for the
  problem to be solved, and has an individual
  behavior.
• When interacting between them and their
  environment, each agent follows its own strategy.
• Within this context, solutions emerge.

15
The origins

• Control Theory Vs. Cognitive psychology
• Theory Control
• Cognitive Psychology
• Classical AI Planning systems

16
Philosophical roots

• Origins in the 18th century.
• Foundation of model control theory laid by James
  Watt.
• Mechanical feedback to control steam engines.
• Cybernetics tried to unify the phenomena of
  control and communication observed in animals and
  machines into a common mathematical model.

17
Agents

• This term is used to characterize, starting from
  primitive biological systems, very different
  kinds of systems.
• Biological ants, bees.
• Movil Robots and air planes.
• Systems that simulate or describe whole human
  societies or organizations such as
• shiping companies
• industrial enterprises

18
A black box agent model
OUTPUT
INPUT
f
Perception
Comunication
19
An agent is internally described through a
function f

• f is a function which takes perception and
  received messages as input and generates output
  in terms of performing actions and sending
  messages.
• The mapping f itself is not directly controlled
  by an external authority the agent is autonomous.

20
This general view of an agent allows its
modelling through

• Biological models
• Based-kowledge models (this kind of models can be
  defined by mental states)
• What makes this models drastically different is
• the nature of the function f which determines the
  agents behaviour.

21
Cognitive Psychology

• Control theory investigates the agent-world
  relationship from a machine oriented perspective.
• The question of how goals and intentions of a
  human agent emerge and how they finally lead to
  the execution of actions that change the state of
  the world, is the subject of cognitive
  psychology, particularly of motivation theory.

22
From Motivation to Action
Resulting motivation tendency
Motivation
Formation of intentions
Decision
Initiation of action
Action
23
Motivational Theory

• The motivation theory study is centered around
  the problem of finding out why an agent performs
  a certain action or reveals a certain behaviour.
  This covers the transition from motivation to
  action where two subprocesses that define two
  basic directions in motivation theory are
  involved.

24

• Formation of intentions how intentions are
  generated from a set of latent motivation
  tendencies.
• Volition and action how the actions of a person
  emerge from its intentions.
• The investigation of reasons, motivations,
  activation, control and duration of human
  behavior goes back at least to Platón and
  Aristóteles. They defined it along 3 categories
  cognition, emotion and motivation.

25

• The main determinant of motivation was situated
  in the human personality a human being is a
  rational creature with a free will.
• In AI, the human needs and goals have been
  structured in a hierarchical way.

26

• Darwin shifted the focus of motivation research
  from a person-centered to a situation-centred
  perspective.
• He established a duality between the human and
  animal behaviors.
• As a consequence, it was found that many of the
  models corresponding to animal behavior are also
  valid for humans.

27

• Another consequence of Darwins theory is that
  human intelligence was viewed as a product of
  evolution rather than a fundamental quality which
  is given to humans exclusively by some higher
  authority.
• Thus, intelligence and learning became a subject
  of sytematic and empirical research.

28

• In the case of AI, hybrid architectures have been
  develpoed to combine both paradigms
  (person-centred and situation-centred).
• Dynamic theory of action (DTA). (Kurt Lewin
  1890-1947).

29
Dynamic Theory of Action

• It is a model explaining the dynamics of change
  of motivation over time.
• The model starts from a set of behavioral
  tendencies which can be compared to the possible
  goals of a person.

30
Dynamic Theory of Action

• For every point in time t and for each behavioral
  tendency b the theory determines a resultant
  action a tendency.
• That is, how strong is b at time t.
• The maximal tendency is called dominating action
  a tendency at time t.

31

• The input for a DTA are an instant t in the
  stream of behavior, and an action tendency which
  is given by a
• motive (person-centered)
• an incentive (situation-centered)
• The dynamics of a DTA is described by means of
  four basic forces
• instigator
• consummator
• inhibitor
• Resistant force.

32

• The output of the DTA is the resulting tendency
  of action for a and tn which is computed as a
  function of the four forces defined above.
• This work is related with Maes Theory
  (agents can have goals), with the BDI
  architecture, and with the control selection of
  the exhibit mechanism of the pedagogical agents
  behaviors.

33
From the point of view of a computer scientist ...

• How can motives and situations be represented and
  recognized?
• How can the influence of motives and situations
  to the basic forces In, Co, Ini, and Re, be put
  into a computational model.
• Can we reduce an agent to a finite set of
  potential behavioral tendencies?

34
Clasical AI Planning systems

• The planning systems are seen as
• a world state
• a goal state and
• a set of operators
• Planning can be looked as a search in a state
  space, and the execution of a plan will result in
  some goal of the agent being achieved.

35
The analogy with the agents theory

• The agent has a symbolic representation of the
  world.
• The state of the world is described by a set of
  propositions that are valid in the world.
• The action effect of the agent in the environment
  are also described by a set of operators, and the
  resulting world state.

36
Reactive-Agents Architectures

• The design of these architectures is strongly
  influenced by behavioral psychology.
• Brooks, Chapman and Agree, Kelabling, Maes,
  Ferber, Arkin
• These kind of agents are kown as
• behaviour-based
• situated or
• reactive

37
Reactive Agents

• The selection-action dynamics for this type of
  system will emerge in response to two basic
  aspects
• the conditions of the environment
• internal objectives of each agent
• Their main characteristics are
• dynamic interaction with the environment
• internal mechanisms that allow working with
  limited resources and incomplete information

38

• The design of reactive architectures is partially
  guided by Simons hypothesis
• the complexity of an agents behavior can be a
  reflection of its opertating environment rather
  than of a complex design.

39

• Brooks thinks that the model of the world is the
  best model for reasoning
• ... and to build reactive systems based on
  perception and action (essence of intelligence)
• Once the essences of being and reaction are
  available, the solutions to the problems of
  behavior, language, expert knowledge and its
  application, and reasoning, become simple.

40
Functionality Vs. Behavior

• From a functional perspective, classical AI views
  an intelligent system as a set of independent
  information processors.
• The subsumption architecture provides an oriented
  descomposition of the activity in this way a set
  of activity (behaviors) producers can be
  identified.
• The behaviors work in parallel, and are tied to
  the real world through perceptions and actions.

41

• An instigator is a force that pushes the action
  tendency for b at time t.
• A consummator is used to weaken the instigating
  force for b over time. This force is only active
  while the behavioral tendency b is active.
• An inhibitor is a force which inhibits the action
  tendency for b at time t.
• A resistant force weakens the inhibitory force
  over time.

42
Present situation of Geothermics in Mexico

• Up to present geothermal resources in Mexico are
  utlized to produce electrical energy
• Some geothermal resources are utlized for
  different purposes
• Turist
• Therapeutic
• Use of the separated waters or the waste heat for
  industrial in mexican geothermal fields.

43

• However exploration and develpoment activities
  are focused on use of geothermal resources.
• The Universities and the CFE (Comisión Federal de
  Elecricidad)

44

• Regional Geothermal assessment in Mexico was
  completed 1987
• When 92 of the whole territory had been covered
• The remining 8 has no geothermal because of its
  tectonically stable location

45
By 1987 ...

• 545 thermal localities had been identified, which
  grouped around 1380 individual hot points
  including
• Hot springs
• Hot water shallow wells
• Hot soils
• Fumaroles, etc.

46

• By 1990, 42 geothermal zones has been located
• In those zones, pre feasabilty studies
  (geology, fluid geochemistry and geophhysics) had
  been conduced in varynig stages.

47

• From 1990 to 1994 detailied geological studies
  were made in the following geothermal zones
• Las tres vírgenes (Baja California Sur)
• Hidrology
• Tectonics
• stratigraphy
• volcanology

48

• El Ceboruco-San Pedro (Nayarit)
• Hidrology
• Tectonics
• volcanology

49
Geothermal Fileds and Geothermal zones under
exploration in Mexico
50
Drilling Activities

• Currently there are 68 geothermal wells,
  representing 104, 859 drilled meters.
• E.g. In the Humeros Geothermal field two deep
  wells were drilled
• There are in Mexico, up to the present, 356 deep
  wells drilled for electrical use of geothermal
  resources. These wells give a total amount of
  715,090 drilled meters.

51

• Currently Mexico has an installed geothermal
  electric capacity of 753 Mwe
• It represents 7 of the overall country production

52
An example

• One of the objectives of artificial intelligence
  refers to the development of systems that ease or
  increase the level of comfort in the daily life
  of humans. Such is the case for tasks with
  permanent focus on the input data in convergent
  methods or systems that help in the
  decision-making process involved in costly
  processes.

53
An example

• In this example we propose a designs of the
  experts decision making process trough the use
  of a cognitive model, and fuzzy sets to model the
  agents reactive deliberative process.

54

• Software system helps human expert in the
  estimation of the static formation temperatures.
• Furthermore, we will present an example based on
  a behavior developed from an expert in the field
  of geothermal sciences.

55

• An attempt to estimate formation temperatures
  from logged temperatures was solved whit this
  methodology based on reactive decision model.

56
Adaptative Behavior

• Autonomy is also known as adaptive behavior and
  it has the capacity to adjust itself to the
  environment conditions
• It is the essence of the intelligence and it is
  the animal ability to fight continuously against
  the world complex, dynamic and unpredictable.

57

• This ability is seeing in terms of flexibility to
  adjust the behavior compendium to the
  contingencies anytime as a product of the
  interaction with the environment.

58
When we use agents to simulate an adaptative
behavior

• Agents can be developed from two perspectives
• knowledge and automatic learning acquisition
• the domain expertise is codified from a human
  expert
• In our study case we design the adaptative
  behavior taken into account the human expertise

59
the design of the representation of dynamical
environment

• could be from two approaches
• the traditional AI considered that the success of
  an intelligent system is closely related with the
  degree of the domain problem, which can be
  treated as a microworld abstraction (symbolic
  processing approaches), that is, at the same
  time, disconnected of the real world.

60

• There exists another group whose design is usally
  bottom - up, it is an etologic design and bears
  in mind the fundamental steps of animal behavior
  (subsymbolic). These approaches also empathizes
  symbol grounding where various behavior modules
  of an agent interact with the environment to
  produce complex behavior.

61

• However this group concedes that achieving
  human-level artificial intelligence might require
  integration of the two approaches.
• In our study case, referring to a simulator
  control, the behavior agent has to be connected
  to the simulator, which represents a dynamic
  environment, modelling the domain expertise to
  the adaptive process. In this case it represents
  a symbolic grounding representation

62
Agents

• Agents continuously perform three functions
• perceptions of the dynamic conditions from the
  environment
• actions that can change the environment
  conditions
• reasoning for interpreting perceptions, solving
  problems, making inferences and taking an action

63
Agents

• Conceptually perception inputs data for the
  reasoning process and the reasoning process
  guides the action
• In some cases the perception can guide the action
  directly

64

• One of the problems in the design of these agents
  is to establish a decision-making process with
  subjective domains
• Natural environments exhibit a great deal of
  structure that a properly designed agent can
  depend upon and even actively exploit
• Strictly talking about the things required to
  achieve an adaptive behavior, a structural
  congruence between the internal dynamic
  mechanisms of an agent and the external
  environment dynamic is needed

65

• As long as this compatibility exists, both the
  environment and the unit act as mutual sources of
  disturbance, release and conditions alteration
• In this case it is a two non-autonomous dynamical
  systems

66

• The agent (the human expert) and the
  environment (the simulator). The design of these
  systems can be seen as a control problem.

67
A control problem

• have two sub-problems
• the state estimation, consisting in the
  evaluation of the environment (perception) and
  the controllers input.
• regulation, consisting in finding an adequate
  response to the environment state (action)

68
The controller consists of

• a function (f) that estimates the environments
  state
• a function that regulates the environments
  response

69
From the perspective of AI

• the agent has the ability to recognize certain
  class of situations, which derive in objectives
  and thus, develop actions that lead to the
  achievement of these objectives

70

• Most of the environments are too complex to be
  described by differential equations
• The behavior of a shipment company of an airport,
  or cognitive processes involving expertise, need
  a kind of symbolic model
• The classic control theory can not deal with
  incomplete information regarding the environment
  in a successful way

71

• In the case of agents, heuristics are use.
• Its use implies a basic difference because the f
  function can be implemented through differential
  equations or symbolic reasoning

72

• A model having an agent and its environment imply
  the existence of two dynamic systems having
  convergent dynamics that is, the value of their
  state variables do not diverge to infinity, but
  eventually converge to a limit set

73

• Figure 1 shows the dynamical systems and the
  variables of our study case. The WELLBORE DATA is
  included in the symbolic model and these
  variables will make the human-expert (autonomous
  agent) reason. In this example the input data
  used by the human-expert of some variables remain
  constant (the mass flow rate during lost of
  circulation and porosity).

74
Dynamic Systems that constitutes the environment
and the autonomous agent
75
Diagram of the data for the obtaining of existing
temperatures
76
Mental model of experts decision
77
Dependency of agents
78
Logged (TReg) and simulated (TSim) temperatures
for the test well. The resulting formation
temperatures (TMod) are also shown
79
Conclusions

• Due to its usefulness and full applicability many
  areas of computer science have rapidly adopted
  this sample and powerful concept
• On AI the introduction of agents is partially due
  to the final deficulties when we try to solve
  problems considering the features of the external
  world or when the agent is involved in a problem
  solving process

80
Conclusions

• The solutions to address these problems can be
  limited and inflexible if there is not a good
  perception of the external world features.
• As a response to this difficulty, the agents
  receive inputs from the environment through
  devices that allow them to perceive the world.
• In response to these inputs, they develop actions
  causing effects on the environment.

81
Conclusions

• In our example we were established two agents
• An autonomous
• Non-autonomous
• This implies a distributed solution to the
  problem, which consists of finding the existing
  temperatures.

82
Conclusions

• These characteristics provide the properties of
  robustness and answer quality to the system.
• The basic reactive behavior design of the agent
  was carried out through located activity that is
  focused on the agents actions and, therefore, on
  its basic behaviors according to the situation,
  moments and environments.

83
Conclusions

• It is fundamental to find the specific
  perceptions that will cause a certain action on a
  present environment.
• To achieve this, a cognitive model that
  represents the experts decision, was developed.
• This model allows the consideration of the
  different situations that can occur in the
  environment, to achieve an emergent response of
  the system.

84
Conclusions

• The behavior has been formalized taking into
  account all the control variables of the process
  
• a) goal type,
• b) knowledge type and
• c) perception and action of each agent.
• This formalization provides an interaction
  between agents with a well-defined interface that
  guarantee a congruent behavior of the muti-agent
  system (environment-agents or agents-agents)

85
Conclusions

• The temperature behavior in the geothermal well
  has been successfully modeled since the
  difference between simulated and logged
  temperatures is inside the human perception.

86
Conclusions

• Finally, this work is an example of a design
  technique proposed for the development of
  multi-agent systems with reactive
  characteristics, which shows the simplicity (with
  respect to previous works) that has been achieved
  through the development of the software that
  controls a dynamic process that involves many
  variables