Agent Architecture-final form 409 by tabindah

VIEWS: 18 PAGES: 7

More Info
									                                       AGENT ARCHITECTURE


                             Dr. S. Srinivasan, Mr. Amit Kumar Goel,Prof B.K Jha
                             Head, Department Of Computer Sciences,PDM Bhadurgarh
                                               dss_dce@gmail.com
                                                 Research Scholar
                                                   BIT-Mesra
                                            goelkamit@rediffmail.com
                                    Department of Management,BITNoida, India
                                              jha_brajendra@hotmail.com


                                                     ABSTRACT

                 The objective of this paper is to provide some rational, structured access to an
                 analysis of cognitive and agent architectures.[4,5] Some familiar architectures have
                 been considered for this preliminary analysis representing a wide range of current
                 architectures in artificial intelligence (AI)[15]. The aim of the paper is to facilitate
                 both an understanding of current architectures and provide insight to the
                 development of future, improved intelligent agent architectures. The main focus is
                 on discussing about various capabilities these architectures possess , various
                 environments upon which these architectures act , and their memory , knowledge
                 representation they use.

                 Keywords: subsumption, soar, atlantis, theo, progigy, homer, icarus, ralph-mea,
                 multiagent system



1   INTRODUCTION                                                 2 TYPE OF AGENT ARCHITECTURE
                                                                 2.1 Subsumption Architecture
          A complete functioning agent, whether                       Subsumption architecture designed by Brooks -
biological, or simulated in software, or implemented             1986 .Complicated Intelligent behavior is split into
in the form of a robot, needs an integrated collection           Simple behavior modules Organized into layers
of diverse but interrelated capabilities, i.e.                   .Behavior is distributed rather than centralized.
architecture. At present, most work in AI and                    Response to stimuli is reflexive -- the perception-
Cognitive Science addresses only components of such              action sequence is not modulated by           cognitive
an architecture (e.g. vision, speech understanding,              deliberation. The agents are organized in a bottom-up
concept formation, rule learning, planning, motor                fashion. Thus, complex behaviors are         fashioned
control, etc.) or mechanisms and forms of                        from the combination of simpler, underlying ones. No
representation and inference (logic engines,                     explicit representation, no explicit working memory,
condition-action rules, neural nets, genetic                     Heuristic with in each layer as long term memory ,
algorithms) which might be used by many                          No explicit problem solving mechanism[1] ,implicit
components. While such studies can make useful                   planning by task decomposition , no learning
contributions it is important to ask, from time to time,         mechanism ;Reactions to environment pre-wired into
how everything can be put together, and that requires            each module ,no reasoning/inference mechanism.
the study of architectures.
     Here we briefly discuss about some familiar                 2.2 Gat’s atlantis architecture
architectures and their features. The study of these                 Integrating planning and reacting in a
architectures provide a better understanding of how              heterogeneous asynchronous architecture for mobile
an intelligent agent should be . [3]                             agents. It consists of Control layer, sequencing layer,
                                                                 deliberative layer. It is hybrid reactive/deliberative
robot architecture.                                        2.9 Soar
                                                           SOAR was designed by Laird, Newell and Rosen
2.3 Theo Architecture                                      bloom. It is the oldest and largest AI development
    The basic idea is Plan-Then-Compile. It means ,        efforts from 1983. Newell made it a developing
integrating learning , planning and     knowledge          Architectures for intelligent agents (1990). SOAR and
representation . It has a self Improving Problem           BDI have a lot of commonalities. SOAR means
solver .                                                   STATE OPERATOR AND RESULT. Soar use all
                                                           capabilities to be handled by an intelligent agent from
2.4 ICARUS                                                 highly routine to open-ended problems. It uses
Specific representation of long term memory .It uses       symbols (it process symbols only ) . The underlying
3 independent asynchronous modules responsible for         SOAR architecture is Symbolic System. SOAR
perception , planning , effecting                          follows Physical symbolic System Hypothesis
2.5 Prodigy                                                (PSSH). It says that Physical symbol system is
Storing the knowledge in a form of first order             necessary and sufficient condition for general
predicate logic (FOPL) called Prodigy Description          intelligence. The ultimate aim is to make a general
language (PDL). It has a modular architecture that         intelligent agent.
stores the knowledge symbolically. Prodigy is an
Architecture of Planning and Learning. Prodigy does        It is based on production system. The production rule
not subscribe to a universal learning method like Soar     is to govern behavior. Production system uses
but rather uses a number of different learning             problem space [13]. Problem space means set of all
mechanisms. These methods allow Prodigy to be used         (possible) states and a set of operators. Operators
as a test-bed for exploring relationships between          transforms a particular state with in problem space to
problem-solving and learning [6].                          another state in the same space .It has initial stats and
                                                           desired state (goal state). Operators are iteratively
2.6 Adaptive Intelligent Systems [2]                       applied to reach goal . The sequence of steps from the
It can do reasoning and interact with other dynamic        initial state to goal state forms the solution or
entities in real time . It has problem solving             behavioral path . It uses chunking (a way of learning)
techniques. When encountering un-expected situation,       and sub goaling. It uses appropriate knowledge
it decides whether to and how to respond. It focuses       (depends on the problem) such as procedural,
attention on the most critical aspects of current          declarative, episodic, semantic, iconic. It employs full
situation. It will operate continuously without            range of problem solving methods. It interacts with
rebooting and also it is able to coordinate with           outside world.
external agent (more or less similar to human being).      SOAR uses full capabilities when it tries to solve
                                                           problems. Two main principle are Functionality and
2.7 Meta Reasoning                                         Performance. It provides integral approach and this
Many ideas in MAX may be traced to Prodigy. It             helps to build general autonomous agents as well as
rule-based forward – chaining engine that operates on      specific intelligent agents for tasks such as natural
productions. It is designed to support to modular          language understanding , analogical reasoning ,
agents. They are used to respond to a dynamic              planning and others.
environment in a timely manner. Modules are
categorized in to behavior and monitor.                    2.10 Tetlon
                                                           It is a problem solver. It uses two memory areas,
2.8 Homer                                                  short-Term memory and long-Term memory. Like
It is not designed for general intelligence. The           human beings, interruption is allowed .It has a feature
underlying philosophy is to synthesize several key         called Execution Cycle which always looks for what
areas of AI to form one complete system, like              to do next.
planning, learning, natural language understanding,
and robotic navigation. HOMER answers questions            2.11 Ralph-Mea
posed by users and carries out instructions given by       It is multiple execution architecture. Like human
users. It is a modular structure. It consists of memory,   being, selecting best one from the environment.
planner, natural language interpreter and generator,       RALPH – MEA uses Execution Architecture (EA) to
reflective processes, plan executer.                       select from one state to best one. It uses the
                                                           following:
1. Condition action                                      Inductive
2. Action utility                                         Learning
3. Goal – based                                              and
4. Decision Theoretic                                     Concept
                                                         Acquisitio
2.12 Entropy Reduction Engine                                 n            Y       Y               Y         Y
It focuses on problems that require planning ,           Abstarctio
scheduling and control . It uses many different               n                Y                   Y
problem solving methods such as problem reduction ,      Explanati
temporal projection , rule-based execution .              on-based
                                                          learning         Y Y             Y       Y         Y
3 CAPABILITIES.                                           Transfer
A complete functioning agent, whether biological, or          of
simulated in software, or implemented in the form of      Learning                   Y
a robot, needs an integrated collection of diverse but    Planning     Y Y Y Y Y Y Y Y Y Y Y
interrelated capabilities. While such studies can make    Problem
useful contributions it is important to ask, from time    Solving              Y       Y       Y Y Y         Y
to time, how everything can be put together, and that    Replannin
requires the study of architectures. [5]. The                 g        Y           Y Y Y Y Y             Y
capabilities are broadly split as learning, planning,
                                                         Support of
problem solving and reasoning.
                                                          Multiple,
In the following table-1A and Table-1B the rows
                                                         Simultane
represent the capabilities and the column corresponds
                                                            -ous
to agents with specific architecture. In the cells ‘Y’
                                                            Goals      Y Y             Y       Y       Y Y
indicates that agent corresponds to the column
                                                             Self
possess the capability represented by the row.
                                                         Reflection        Y Y Y Y Y Y
                                                           Meta-
Table 1: Capabilities of Agents
                                                         Reasoning         Y Y         Y Y         Y     Y
                                                           Expert
                      Meta Reasoning Architecture
                      Adaptive Intelligence System




                       Entropy Reduction Engine




                                                           System
                                                         Capability                    Y           Y
                                                         Inductive
                             Subsumption




                              Ralph-Mea
         Capability


                                Atlantis




                                                             and
                                Prodigy




                                Homer

                                 Tetlon
                                 Icarus
                                  Theo




                                  Soar




                                                         deductive
                                                         Reasoning
                                                         Prediction        Y Y Y Y                       Y Y
                                                           Query
                                                         Answerin
                                                            g and
                                                         providing
   Single                                                Explanati
  Learning                                                 ons for
  Method                         Y       Y Y             Decisions         Y           Y       Y
   Multi-                                                Navigatio
  Method                                                     nal
  Learning                 Y Y       Y     Y       Y     Strategies   Y Y Y        Y Y         Y
  Caching                  Y               Y               Natural
  Learning                                               Language
  by Inst.                   Y             Y             Understan
  Learning                                                  ding                               Y Y
  by Expt.                   Y       Y                    Learning
  Learning                                                    by
     by                                                  Perceptio
  Analogy                    Y             Y                  n       Y Y Y            Y       Y Y           Y
  Support                                                  environment represented by the row.
    for
 Inaccurate                                                Table 2: Environment of Agents
  sensing




                                                                         Meta Reasoning Architecture
                                                                         Adaptive Intelligence System
                  Y             Y                  Y




                                                                          Entropy Reduction Engine
  Robotic
   Tasks       Y Y Y            Y      Y Y         Y




                                                                                Subsumption




                                                                                 Ralph-Mea
 Real-time




                                                                                   Atlantis

                                                                                   Prodigy




                                                                                   Homer

                                                                                    Tetlon
                                                                                    Icarus
 Execution




                                                                                     Theo




                                                                                     Soar
  Focused
  Behavior
     and
 Processin
     g/
  Selective
 Attentions    Y Y Y Y Y Y Y Y Y                   Y          Static
    Goal                                                   Environme
 Reconstru                                                      nt                Y                  Y
    ction                                      Y            Dynamic
  Respondi                                                 Environme
      ng                                                        nt       Y Y Y        Y Y Y Y Y              Y Y
 Intelligent                                               Consistent
    ly to                                                  Environme
 Interrupts                                                     nt             Y Y                   Y
     and                                                   Simulated
  Failures        Y      Y Y Y Y Y Y               Y
                                                           Environme
  Human-                                                        nt                Y          Y Y Y Y Y
 like Math
                                                              Real
 Capability                                Y
                                                             World
                                                           Environme
4 ENVIRONMENT                                                   nt       Y Y Y        Y Y            Y
The world itself is a very complex place for                Complex
architecture agents to act in. Most architecture are       Environme
designed to only deal with fractions of the total
                                                                nt       Y Y Y        Y Y Y          Y
possible environmental complexity by acting in
                                                           Knowledg
particular domains. For example, some architecture
                                                             e-Rich
assume that the world is static and that the only things
                                                           Environme
that change in the world are via an agent's actions.
                                                                nt             Y Y                   Y
Other architectures may operate in dynamic
                                                           Input-Rich
environments but require that world be consistent or
                                                           Environme
predictable.
                                                                nt                       Y
An agent is anything that can be viewed as perceiving
                                                             Limited
its environment through sensors and acting upon that
                                                           Resources           Y Y       Y       Y
environment through actuator [2]. A Vacuum cleaner
agent is a reactive agent which means it responds           Complex
when it senses dirt on the floor . Where as a Taxi         Knowledg
Driver agent should behave in a different way since             e                 Y Y Y          Y       Y
its environment is entirely different. It is a Goal            Un-
directed agent .[ 8] It has to keep all the percepts it    predictable
receives in its knowledge base .                                         Y            Y Y        Y Y         Y Y
In this section various environments with which the        Asynchron
agents have to work are discussed. In table-2 , Rows          ous        Y               Y
indicate the environments and the column indicates         Concurren
the particular architecture . Y in the cell means that         t
the architecture corresponding to column uses the                        Y Y             Y
 Varying




                                                                                        Meta Reasoning Architecture
                                                                                        Adaptive Intelligence System
 Priority                   Y Y Y                Y




                                                                                         Entropy Reduction Engine
                                                                 Memory , Knowledge &
 Limited
Response




                                                                    Representation
                                                                                               Subsumption




                                                                                                Ralph-Mea
  Time                      Y Y                  Y Y




                                                                                                  Atlantis

                                                                                                  Prodigy




                                                                                                  Homer

                                                                                                   Tetlon
                                                                                                   Icarus
 Multiple




                                                                                                    Theo




                                                                                                    Soar
  Tasks                        Y       Y
Supervisor                     Y           Y

5   MEMORY,    KNOWLEDGE                         AND
KNOWLEDGE REPRESENTATION

Here the focus is to study Memory , Knowledge and          Forward &
Knowledge representation are handled by Agent              Back-ward
architecture . The aim of the paper is to understand        Chaining                                Y Y             Y
the various Knowledge Representations that agents           Impasse-
should adapt generally . Also because of the design of       driven
the architecture of the agents that are taken in to          Control                            Y     Y        Y Y Y
consideration, the representations vary from one agent        Serial
to another [11,12].                                        Processing                               Y Y       Y Y Y     Y
                                                            Parrellel
A Vacuum cleaner agent is a reactive agent which           Processing                                 Y Y           Y   Y Y
means , it responds when it senses dirt on the floor .    Asynchronou
Where as a Taxi Driver agent should behave in a           s Processing                      Y             Y                 Y
different way since its environment is entirely           Interruptible
different. It is a Goal directed agent. [ 8 ] it has to    Processing                                 Y Y      Y Y
keep all the percepts it receives in its knowledge base    Open-Loop
. The architectures of these two typical examples          processing                                 Y                     Y
should certainly differ greatly . Agent properties          Closed -
identify and entail the techniques and methods that           Loop
were used to realize a particular architecture or          Processing                                 Y
architectural component. For example, most                Hierarchical
architecture includes some sort of memory. An agent       Organization                  Y Y           Y Y           Y   Y
is said to be a knowledge-level system [14] when it         Modular
rationally brings to bear all its knowledge onto every    Organization                  Y           Y Y Y Y Y           Y Y
problem it attempts to solve. Thus, knowledge is the        Symbolic
medium of transaction at the knowledge level and the      World Model                       Y Y Y Y Y Y Y Y             Y Y
behavioral law is the principle of maximum                 Size of the
rationality. Agent properties characterize the memory:    Knowledge –
Is the memory declarative, procedural, and episodic?          Base                              Y         Y         Y
Are there size limitations? Is memory uniformly            Glass Box
accessed? Is it uniformly organized? These properties       approach                            Y Y           Y Y
have been shown in the form a table .                      Black Box
                                                            Approach                        Y             Y    Y Y      Y
In the Table 3 , Rows indicate the Memory ,                Declarative
Knowledge and knowledge representation and the            Representati
column indicates the particular architecture . Y in the        on                                   Y Y Y Y             Y Y
cell means that the architecture corresponding to          Procedural
column has the type of memory or knowledge and            Representati
knowledge representation .                                     on                               Y     Y             Y       Y
Table 3: Memory Knowledge representation                     Global
                                                          Representati
                                                               on                               Y Y Y Y Y Y Y Y Y Y
   Uniform




                                                                                Meta Reasoning Architecture
                                                                                Adaptive Intelligence System
   Access to




                                                                                 Entropy Reduction Engine
  Knowledge                           Y Y
  Knowledge




                                                                                       Subsumption




                                                                                        Ralph-Mea
 Consistency            Y       Y     Y




                                                                   Evaluation


                                                                                          Atlantis

                                                                                          Prodigy




                                                                                          Homer

                                                                                           Tetlon
                                                                                           Icarus
 Homogenous




                                                                                            Theo




                                                                                            Soar
  (Uniform)
  Knowledge
 Representati
       on               Y Y Y Y Y
 Hetrogenous
  Knowledge
 Representati                                              Concept
       on                                                       s
  No-Explicit                                               Internal
 Representati                                              Architectu
       on         Y                                            re               Y                  Y
   Associate                                                 Social
   Memory                                                  architectur
                                                                e                                  Y
   Episodic
                                                           Communi
  Knowledge                               Y                  cation
     Meta-
  Knowledge             Y             Y Y Y Y              Autonomy                                Y
  First-Order
                                                              Pro-
     Logic                                                  Activity                               Y
 Representati                                              Distributio
       on                   Y                                   n
  Strips like                                               Notation
   operators
                                                            Usability           Y
 representatio
                                                           Expressiv
        n                   Y         Y Y                    eness
 Frame - Like                                              Refineme
 Representati                                                  nt
       on               Y             Y Y                  Dependan
   Network                                                    cy of
 representatio                                              Models
        n                       Y Y                        Traceabilit
                                                                y
6 EVALUATIONS                                                Clear
                                                           definition                              Y
Based on the concepts, notation, process and               Modularit
pragmatics the architectures are evaluated. These are           y               Y                  Y
the criteria and scales for evaluating architectures. In
table-4 the column represents selected architectures        Process                                Y
and rows represent the criteria considered for             Coverage
                                                               of
evaluation. Y stands for the architecture being
                                                           workflows
evaluated against particular criteria representing the
                                                           Managem
row . But most of the cell is blank since no proper            ent
methodologies are available. Also due to the lack of       Complexit
analytical criteria, the cost of demonstrations and             y
varying specifications among different architectures,      Properties
developing evaluation methods is a challenge [7, 10].      of process
                                                           Pragmatic
Table 4: Evaluations                                            s
   Tool                                                         Computers and Thought, ed. Feigenbaum and
  Support                                                       Feldman. McGraw-Hill, New York.
 Connectiv                                                  [4] Simon, H. (1991). Cognitive Architectures in a
    ity                                                         rational analysis: comment. In K. VanLehn (ed.),
 Document                                                       Architectures for Intelligence, pp. 25-39,
   ation
                                                                Lawrence Erlbaum Associates, Hillsdale, N.J.
 Usage in
  Projects    Y                                             [5] Newell, A. (1990). Unified Theories of
                                                                Cognition. Harvard University Press. Cambridge,
                                                                Massachusetts.
                                                            [6] Elavine Rich , Kevin Knight and Shivshankar ,
CONCLUSION
The problem of AI is to describe and build agents that          Artificial Intelligence by The Mcgraw Hill
receive percepts from the environment and perform               Publishing Company limited .
actions, and each such agent is implemented by a            [7] Simon, H. (1962). The architecture of
function that maps percepts to actions. It explains the         complexity. Proceedings of the American
role of learning as extending the reach of the designer         Philosophical Society, 26, 467-482.
into unknown environments, and shows how it                 [8] Artificial Intelligence: Modern Approach by
constrains agent design, favoring explicit knowledge            Stuart J. Russell , Peter Norvig , Prentice Hall
representation and reasoning . It analyzes basic                Series in Artificial Intelligence.
techniques for addressing complexity . This paper has       [9] Kurt Vanlehn , Rule-Learning Events in the
tackled the question how a developer can choose                 acquisition of a complex skill , Journal article by,
among the many development options when                         Journal of the Learning Science ,vol 8 , 1999
implementing an agent application. One key aspect           [10] Philip E. Agre , Hierarchichy and History in
here is to understand that agent technology currently           Simon’s “Architecture of Complexity “ Journal
o ers many problem specific solutions that address               of Learning Science 12(3) , 2003
only certain types of application domains. We argue         [11] Levesque H , Brachman R , A fundamental trade
that one important foundation for making accurate               off in Knowledge representation and reasoning ,
choices is the availability of well-defined and                  1998
comparable surveys and evaluations of artifacts such        [12] Vincent C.Muller , Is there a future for AI
as environment and capabilities. Therefore, we have             without representation , springer Verlag , 2007
in a tabular form for evaluating many kinds of              [13] A.S Maida , A Uniform architecture for rule-
Architectures with respect to capability, environments          based meta-reasoning and representation :- case
, memory and knowledge representation . .In future              study . , IEEE Computer Society Press.
work we want to employ on these tables to study             [14] Newell, A. (1982). The knowledge level.
Multi-Agent System Technology. The idea is to                   Artificial Intelligence. 18(1), 87-127.
Integrate state-of-the art AI techniques into intelligent   [15] http://ai.eecs.umich.edu/cogarch0/
agent designs, using examples from simple, reactive
agents to full knowledge-based agents with natural
language capabilities and so on . This leads to the
study of Multi-Agent systems and its applications. In
depth analysis of various Agent architectures is to
build a Multi Agent System that will be suitable for
our future work on Supply Chain Management.

REFERENCES
[1] Fikes, R., Nilsson, N. (1971).strips: A new
    approach to the application of theorem proving to
    problem solving. Artificial Intelligence, 2, pp.
    189-203.
[2] Firby, R. (1989). Adaptive execution in complex
    dynamic worlds. Ph.D. thesis. Department of
    Computer Science. Yale University. New Haven,
    Connecticut.
[3] Newell, A. and Simon, H. (1963). GPS: A
    program that simulates human thought. In

								
To top