PARADIGMS OF AI %---------------------------------------------------------------------------- What kind of science is AI? According to Susan Josephson, a philosopher of science, different views of AI as a science are characterized by the role played by computer programs: AI as an Information Science: representation, organization, storage, retrieval, use, display, and acquisition of information AI as a Cognitive Science: description, analysis, modeling, explanation, and prediction of animal behavior (mainly human behavior) AI as a Design Science: interpretation of problems, analysis of knowledge available in the world, design and construction of intelligent (autonomous and interactive) systems for solving the problems %---------------------------------------------------------------------------- What is the relation between AI and computer science? According to Margaret Boden, a philosopher of mind, ``AI is not the study of computers, but of intelligence in thought and action. Computers are its tools, because its theories are expressed as computer programs that enable machines to do things that would require intelligence if done by people.'' In this sense, AI may thought of as applied computer science. In practice, AI makes liberal use of concepts, methods and tools from automata theory, algorithm theory, computer architecture, computer systems, database systems, programming languages, and software engineering. %---------------------------------------------------------------------------- Scientific Paradigms According to Thomas Kuhn (a philosopher of science), all sciences pass through different phases of development which are characterized by different paradigms. The `normal' phase of a science is characterized by an established paradigm that almost all researchers in the discipline accept (for example, the Newtonian paradigm of classical physics). The `revolutionary' phase of a science is characterized by a shift from the old paradigm to a new one (for example, the shift from the Newtonian paradigm of classical physics to the quantum paradigm of modern physics at the turn of the century). The `pre-scientific' phase is charaecterized by multiple and competing paradigms. In this sense, AI is still a `pre-science': as yet, there is no established, commonly accepted AI paradim. In practice, AI researchers are split into several competing schools of thought that are pursuing competing paradigms. %---------------------------------------------------------------------------- AI Paradigms An AI paradigm can be characterized by the constraints it accepts, the assumptions and hypotheses it makes, and the cannonical examples of its success or failure. Some examples of AI paradigms: Functionalism, teleological models, goal/task-directed processing Conceptual information processing, internal representations, mental models Symbolic representations, logic, logical methods Numerical representations, numerical methods, Probalistic representations, statistics, statistical methods Situated action, reactive control Connectionism (neural networks, associative memories, subsymbolic processing) Dynamical systems (physical systems, systems analysis and engineering) Evolutionary systems (biological systems, genetic algorithms) Distributed systems, multi-agents %---------------------------------------------------------------------------- Assumptions of AI Every science makes some assumptions, posits some hypotheses. All AI researchers subscribe to one assumption. Assumption 1: Computation The fundamental working assumption, or "central dogma", of AI is this: What the brain does may be thought of at some level as a kind of computation. --- Charniak and McDermott Operationally: Computation implies processing of information Issues: what kinds of information? what kinds of processing? AI researchers in the `functional' paradigm subscribe to a second assumption as well. Assumption 2: Goals/Tasks Intelligent behavior is goal/task-oriented (or at least it is useful to ascribe goals to intelligent agents and analyze their behavior in terms of the goals). Issues: What kinds of goals/tasks? How is the information processing directed by the goals/tasks AI researchers in the `conceptual' paradigm subscribe to a third assumption as well. Assumption 3: Knowledge/Experience Knowledge of the world and experiences in the world play a central role in cognition. Issues: What kinds of knowledge? Experience? How is this knowledge acquired? How is this knowledge used in the service of goals/tasks? AI researchers in the `symbolic' paradigm subscribe to a fourth assumption, which Newell and Simon (two of the founding fathers of AI) called the Physical Symbol System Hypothesis. Assumption 4: Physical Symbol Systems Hypothesis: A physical symbol system has the necessary and sufficient means for general intelligent action. Operationally: a. Symbols are patterns b. Symbols represent the real world. c. Symbols may point to other symbols d. Symbols are instantiated in physical systems. e. Symbols can be combined to make complex symbolic structures f. Symbolic structures can be manipulated by processes to produce new structures (or expressions) Issues: What kind of symbolic structures? What kinds of manipulations can be done on the structures? What kinds of processes do the manipulations Much of our study of AI will focus on this computational-functional-conceptual-symbolic paradigm of AI. %---------------------------------------------------------------------------- What is an AI model? The assumptions and hypotheses of the C-F-C-S paradigm directly lead to a characterization of an AI model: 1. The model is computational. 2. The behavior of the model is goal-directed; the model explicitly specifies the goals. 3. The model is process-oriented; the model explicitly specifies the mechanisms by which the goals are achieved. 4. The process is knowledge-based (and/or experience-based); the model explicitly specifies the types of knowledge it uses. 5. The knowledge is represented symbolically. %---------------------------------------------------------------------------- Constraints on AI Models In addition, AI models must satisfy some other constraints. Different researchers within the C-F-C-S paradigm differ in the constraints they impose on their models. Constraints from computer science: One constraint generally accepted by almost all AI researchers is that AI models must be instantiable in computer programs. Note that computer program is only an instantiation of an AI model, not the model itself. Some of other constraints are: Constraints from psychology: AI models must explain psychological data; this leads to research on cognitive science Constraints from biology: AI models must account for the architecture of the brain; this leads to research on connectionist models Constraints from mathematics: AI models must be well-defined; this leads to research on formalisms such as logic and probability. Constraints from engineering: AI systems must be able to solve practical problems (or at least help humans solve them); this leads to research on AI systems design. Constraints from the physical world: AI artifacts must be able to operate (act, see) in the physical world; this leads to research on robotics and vision. Constraints from the social world: AI agents must be able to interact with other agents; this leads to research on distributed AI. In practice, most AI researchers accept some non-null subset of the these constraints in addition to the computational constraint. %---------------------------------------------------------------------------- How do we build an AI model? 1. Define the task: What is the intelligent activity that the system will perform? 2. Define the domain: What is the domain in which the system will operate? 3. Define the knowledge: a. What does the system need to know in order to perform the task? b. What kinds of knowledge are available in the given domain? 4. Design the representations: How will that knowledge be encoded in inside the machine? 5. Design the memory: a. How will the system know which piece of knowledge to use when b. How will it access it? 6. Design the process: How will the system use the knowledge to perform the given task? %---------------------------------------------------------------------------- Some difficulties in building an AI model The computation problem AI tasks are generally computationally hard, often computationally intractable AI systems have only limited computational resources bounds on processing bounds on memory The goal problem An AI system can have multiple, interacting goals Goals are sometimes hard to chaarcterize accurately Goals can change over time The knowledge problem Knowledge is voluminous. Knowledge is hard to characterize accurately. Knowledge is constantly changing. The more knowledge you need, the harder it is to: Anticipate all the knowledge that will be needed. Program it all in by hand. Design algorithms that will find the right piece of knowledge at the right time. The behavior problem It is often hard to analyze the behavior of an AI system More complex is the system, more complex is its behavior More complex is the behavior, harder it is to decide what in the system is making the behaving in a particular way %---------------------------------------------------------------------------- Four Basic AI Models of Intelligent Agents (from Russell and Norvig): 1. Situated action, reactive control: Direct mapping of perceptions of the world into actions on it; no internal representations (the world is its own best representation); no memory (actions depend only on the current state of the agent and the world). Advantages: rapid response; Issues: local minina. 2. Memory and Experience: memory of past states of the agent and interactions with the world guides actions (in addition to perceptions); but still little deliberative reasoning. Advantages: takes history into account; avoids past failures; Issues: memory organization, dynamic worlds. 3. Knowledge and Models: internal representations and deliberative reasoning guide actions on the world (in addition to perceptions and memory). Advantages: global view; Issues: knowledge representation, control of processing, computational efficiency. 4. Utility: abstract notions of utility guide actions on the world (in addition to perceptions, memory, and deliberation). Advantages: goes beyond goals; Issues: decision-making, computational efficiency. All four agent models admit multi-agent societies. %---------------------------------------------------------------------------- Used and adapted with permission: Copyright (c) Ashok Goel 1994-96 College of Computing Georgia Institute of Technology, Atlanta, Georgia 30332-0280 goel@cc.gatech.edu %----------------------------------------------------------------------------