Cognitive Architecture B. Kiran Maruthi 09005047 M. Sonika 09005054 D. V. Ramana 09005059
OUTLINE What is Cognitive Architecture? Plausibility of Cognitive Architectures TypeIdentity Theory Functionalism History of Cognitive Architecture General Characteristics Consciousness Unified Theory of Cognition SOAR – case study
Intelligent Agents Entities which observe through sensors and act upon the environment using actuators and direct their activity towards achieving goals.
What is Cognitive Architecture? Blueprint for intelligent agents. It proposes (artificial) computational processes that act like cognitive systems (human) An approach that attempts to model behavioral as well as structural properties of the modeled system. Aim : to model systems that accounts for the whole of cognition, i.e., systems with Artificial Consciousness – which can not only respond but also think, perceive and believe like a human !
Artificial Consciousness Artificial Consciousness is broadly classified as access and phenomenal consciousness. Brain processes neural impulses from the eyes and determines that this image is physically unstable – pattern recognizability. What about pain, anger, motivation, attention, feeling of relevance, modeling other peoples intentions, anticipating consequences of alternative actions, or inventing ?
Plausibility of Artificial Consciousness A view skeptical of AC is held by typeidentity theorists “consciousness can only be realized in particular physical systems because consciousness has properties that necessarily depend on physical constitution” However, for functionalists, “any system that can instantiate the same pattern of causal roles, regardless of physical constitution, will instantiate the same mental states, including consciousness” Along these lines, some theorists have proposed that consciousness can be realized in properly designed and programmed computers.
TypeIdentity Theory The mental events can be grouped into types and associated with types of physical events in the brain. For example, mental event pain results in physical event in the brain (like Cfiber firings) We have two totally different versions of typeidentity theory based on the definition of “what kind of identity” is associated with mental and physical events. Ullin Place (1956) – Compositional Identity Feigl (1957) and Smart (1959) – Referential Identity
Compositional TypeIdentity Theory U.T.Places notion of identity is described as a relation of composition. Every mental process is composed of a set of physical sensations to which it reacts. But can we associate them based purely on composition? "lightning is an electrical discharge" is true.
Referential TypeIdentity Theory For Feigl and Smart, the identity was to be interpreted as the identity between the referents of two descriptions which referred to the same thing. “the morning star” and “the evening star” are identical in the sense that both of them refer to the Venus. Sensations and brain processes do indeed mean different things but they refer to the same physical phenomenon. This is called as The Fregean distinction Conclusion : All of the versions share the central idea that the mind is identical to something physical.
Multiple Realizability Objections to the type identity theory Hilary Putnam popularized it in late 1960s. It states that the same mental property, state, or event can be implemented by different physical properties, states or events.
Putnams Formulation Do all organisms have the same brain structures? Clearly not ! Pain corresponds to completely different physical states and yet they all experience the same mental state of "being in pain." Should robots be considered a priori incable of expereincing pain just because they did not posses the same neurochemistry as humans? Putnam concluded that typeidentity is making an implausible conjecture.
Functionalism Core idea is that mental states are constituted solely by their functional role They are causal relations to other mental states, sensory inputs, and behavioral outputs. Brains are physical devices with neural substrate that perform computations on inputs which produce behaviours. According to this theory it is possible to build silicon based devices which are functionally isomorphic to the humans as long as system performs appropriate functions.
Variations of Functionalism Machine State functionalism – Hilary Putnam Mental state is like automaton state of a Turing Machine. Each state can be defined exclusively in terms of its relations to the other states as well as inputs and outputs. Being in pain is the state which disposes one to cry "ouch"!
Variations of FunctionalismCont... Psycho functionalism – Jerry Fordor The role of mental states, such as belief and desire, is determined by the functional or causal role that is designated for them within our best scientific psychological theory. If some new mental state from folk psychology comes,it is considered nonexistent as it has no fundamental role in cognitive psychological explaination. Some theoretical cognitive psychological states which are necessary for explaination of human behaviour but are not
Quick Question What difference does the colour RED make?
Qualia From the Latin, meaning "what kind". refers to the subjective qualities of sensory perception and the feeling they generate. Qualia is not only the “redness” of red, but the way that redness makes us feel. Qualia are, in essence, our own unique and personal perceptions of our environment.
Marys thought experiment Frank Jackson offers the knowledge argument for qualia. Mary, the colour scientist knows all the physical facts about colour and the experience of colour with other people. Confined from birth to a room that is black and white. When she is allowed to leave the room, it must be admitted that she learns something about the colour red the first time she sees it — specifically, she learns “what it is like” to see that colour. This attacks the knowledge completeness of functionalism.
Ability Hypothesis Nemirow claims that "knowing what an experience is like is the same as knowing how to imagine having the experience". He argues that Mary only obtained the ability to do something, not the knowledge of something new. Mary gained an ability to "remember, imagine and recognize." Knowing what its like to see red is merely a sort of practical knowledge, a “knowing how” (to imagine, remember, or reidentify, a certain type of experience)
Functional IsomorphismPutnam defined the concept of functional isomorphism as : Two systems are functionally isomorphic if there is a correspondence between the states of one and the states of the other that preserves functional relations.
Presently...Functionalism is widely accepted and research to develop cognitive robots is on!
Cognitive Architecture Using Putnams Multiple Realizability formulation and functionalism, David Chalmers in late 1960s suggested the possibilty of mechanisms and structures that underlie Cognition : processors that manipulate data memories that hold knowledge and interfaces that interact with an environment.
History of Cognitive Architecture19692000(time line)
• GPS (Ernst & Newell, 1969) Means-ends analysis, recursive subgoals1970 • ACT (Anderson, 1976) Human semantic memory • CAPS (Thibadeau, Just, Carpenter) Production system for modeling reading1975 • Soar (Laird, & Newell, 1983) Multi-method problem solving, production systems, and problem spaces • Theo (Mitchell et al., 1985) Frames, backward chaining, and EBL1980 • PRS (Georgeff & Lansky, 1986) Procedural reasoning & problem solving • BB1/AIS (Hayes-Roth & Hewitt 1988) Blackboard architecture, meta-level control1985 • Prodigy (Minton et al., 1989) Means-ends analysis, planning and EBL • MAX (Kuokka, 1991) Meta-level reasoning for planning and learning1990 • Icarus (Langley, McKusick, & Allen,1991) Concept learning, planning, and learning • 3T (Gat, 1991) Integrated reactivity, deliberation, and planning1995 • CIRCA (Musliner, Durfee, & Shin, 1993) Real-time performance integrated with planning • AIS (Hayes-Roth 1995) Blackboard architecture, dynamic environment2000 • EPIC (Kieras & Meyer, 1997) Models of human perception, action, and reasoning • APEX (Freed et al., 1998) Model humans to support human computer designs
Characteristics Holism, e.g. Unified theory of cognition The architecture often tries to reproduce the behavior of the modelled system (human), in a way that timely behavior (reaction times) of both are comparable Other cognitive limitations are often modeled as well Robust behavior Parameter – free Artificially Conscious
Artificial ConsciousnessThe functions of consciousness suggested by Bernard Baars : Definition and Context Setting Adaptation and Learning Anticipation Function Prioritizing and AccessControl Decisionmaking or Executive Function Analogyforming Function Metacognitive and Selfmonitoring Function Autoprogramming and Selfmaintenance Function Definitional and Contextsetting Function.
Learning Reaction time for consecutive readings? Human improvement via Practise
Anticipation Machine needs flexible, realtime components that predict worlds. A conscious machine should make coherent predictions and plans, for environments that may change. Executed only when appropriate to simulate and control the real world. Significant research on role of consciousness in cognitive models. Examples : CLARION, OpenCog
Unified Theory of Cognition Book written by Allen Newell Newells goal : To define the architecture of human cognition, which is the way that humans process information. This architecture must explain how we react to stimuli, exhibit goal directed behavior,acquire rational goals, represent knowledge, and learn.
Newells Cognitive Model Newell introduces Soar, an architecture for general cognition. Soar is the first problem solver to create its own subgoals and learn continuously from its own experience. Soar has the ability to operate within the realtime constraints of intelligent behavior, such as immediate response and itemrecognition tasks.
Soar What is Soar? History of Soar Architecture of Soar Evolution of Soar and present version
What is Soar? Soar is a symbolic cognitive architecture. An AI programming language. It provides a (cognitive) architectural framework, within which you can construct cognitive models. It can be considered as an integrated architecture for knowledgebased problem solving, learning, and interaction with external environments.
History Created by John Laird, Allen Newell, and Paul Rosenbloom at Carnegie Mellon University in 1983.John Laird Allen Newell Paul Rosenbloom
Its Soar not SOAR ! Historically, Soar stood for State, Operator And Result, because all problem solving in Soar is regarded as a search through a problem space in which you apply an operator to a state to get a result. Over time, the community no longer regarded Soar as an acronym: this is why it is no longer written in upper case
Screenshot – Soar Debugger
Problem Spaces Soar represents all tasks as collections of problem spaces. Problem spaces are made up of a set of states and operators that manipulate the states. Soar begins work on a task by choosing a problem space, then an initial state in the space. Soar represents the goal of the task as some final state in the problem space.
Structure of Soar Soar can be divided into 3 levels : Memory Level Decision Level Goal Level
Memory Level A general intelligence requires a memory with a large capacity for the storage of knowledge. A variety of types of knowledge must be stored, including : declarative knowledge procedural knowledge episodic knowledge
Longterm Production Memory All of Soars longterm knowledge is stored in a single production memory. Each production is a conditionaction structure that performs its actions when its conditions are met. Memory access consists of the execution of these productions. During the execution of a production, variables in its actions are instantiated with value.
Working Memory The result of memory access is the retrieval of information into a global working memory. It is the temporary memory that contains all of Soars shortterm processing context. It has 3 components : The context stack specifies the hierarchy of active goals, problem spaces, states and operators objects, such as goals and states (and their subobjects) preferences that encode the procedural searchcontrol knowledge
Preferences There is one special type of working memory structure “the preference” Preferences encode control knowledge about the acceptability and desirability of actions. Acceptability preferences determine which actions should be considered as candidates. Desirability preferences define a partial ordering on the candidate actions.
Decision Level The decision level is based on the memory level plus an architecturally provided, fixed, decision procedure. The decision level proceeds in a two phase elaboratedecide cycle. During elaboration, the memory is accessed repeatedly, in parallel, until quiescence is reached; that is, until no more productions can execute. This results in the retrieval into working memory of all of the accessible knowledge that is relevant to the current decision. After quiescence has occurred, the decision procedure selects one of the retrieved actions based on the preferences that were retrieved into working memory.
Goal Level A general intelligence must be able to set and work towards goals.This level is based on the decision level. Goals are set whenever a decision cannot be made; that is, when the decision procedure reaches an impasse. Impasses occur when there are no alternatives that can be selected (nochange and rejection impasses) or when there are multiple alternatives that can be selected, but insufficient discriminating preferences exist to allow a choice to be made among them (tie and conflict impasses).
Impasse Resolution Whenever an impasse occurs, the architecture generates the goal of resolving the impasse which becomes the subgoal. Along with this goal, a new performance context is created. The creation of a new context allows decisions to continue to be made in the service of achieving the goal of “resolving the impasse”. A stack of impasses is possible. The original goal is resumed after all the impasse stack is
Learning through Chunking In addition to all above levels, a general intelligence requires the ability to learn. All learning occurs by the acquisition of chunks productions that summarize the problem solving that occurs in subgoals, a mechanism called “Chunking” The actions of a chunk represent the knowledge generated during the subgoal; that is, the results of the subgoal.
Evolution of Soar YEAR VERSION IMPLEMENTED IN1982 Soar 1 Lisp1983 Soar 2 Lisp/OPS51984 Soar 31986 Soar 41989 Soar 51992 Soar 6 C1996 Soar 7 Tcl/tk1999 Soar 8 SGIO
Soar 9 : Interesting Developement Unifying Cognitive Functions and Emotional Appraisal The functional and computational role of emotion is open to debate. Appraisal theory is the idea that emotions are extracted from our evaluations (appraisals) of events that cause specific reactions in different people. The main controversy surrounding these theories argues that emotions cannot happen without physiological arousal.
Appraisals Detector This theory proposes that an agent continually evaluates a situation and that evaluation leads to emotion. The evaluation is hypothesized to take place along multiple dimensions, such as goal relevance goal conduciveness causality and control These dimensions are exactly what an intelligent agent needs to compute as it pursues its goals while interacting with an environment.
Conclusion This ﬁeld still has far to travel before we understand fully the space of cognitive architectures and the principles that underlie their successful design and utilization. However, we now have over two decades’ experience with constructing and using a variety such architectures for a wide range of problems, along with a number of challenges that have arisen in this pursuit. If the scenery revealed by these initial steps are any indication, the journey ahead promises even more interesting and intriguing sites and attractions.
Soar 9 : Appraisal Detector
References1) SOAR : An Architecture for General Intelligence, John E. Laird, Allen Newell, Paul S. Rosenbloom,1986.2) A preliminary analysis of the Soar architecture as a basis for general intelligence, John E. Laird, Allen Newell, Paul S. Rosenbloom, 1989.3) http://en.wikipedia.org/wiki/Cognitive_architecture4) http://cs.gmu.edu/~eclab/research.html5) http://en.wikipedia.org/wiki/Unified_theory_of_cognition6) http://cll.stanford.edu/research/ongoing/icarus/7) http://en.wikipedia.org/wiki/Artificial_consciousness8) http://plato.stanford.edu/entries/functionalism/
References9) A Survey of Cognitive Architectures, David E. Kieras, University of Michigan .10) Connectionism and Cognitive Architecture : A Critical Analysis, Jerry A. Fodor and Zenon W. Pylyshyn, Rutgers Center for Cognitive Science, Rutgers University, New Brunswick, NJ.11) Human Cognitive Architecture, John Sweller, University of New South Wales, Sydney, Australia.12) http://cogarch.org/index.php/Soar/Architecture13) http://code.google.com/p/soar/wiki/Documentation
References14) A Gentle Introduction to Soar : An Architecture for Human Cognition : 2006 Update, Jill Fain Lehman, John Laird,Paul Rosenbloom.15) http://sitemaker.umich.edu/soar/home