Lecture 1 The Intelligent Agent Framework Friday 22 August 2003 William H. Hsu, KSU http://www.kddresearch.org http://www.cis.ksu.edu/~bhsu Reading for Next Class: Chapter 2, Russell and Norvig
Lecture Outline Today s Reading: Chapter 2, Russell and Norvig Intelligent Agent (IA) Design Shared requirements, characteristics of IAs Methodologies Software agents Reactivity vs. state Knowledge, inference, and uncertainty Intelligent Agent Frameworks Reactive With state Goal-based Utility-based Thursday: Problem Solving and Search State space search handout (Winston) Search handout (Ginsberg)
Why Study Artificial Intelligence? New Computational Capabilities Advances in uncertain reasoning, knowledge representations Learning to act: robot planning, control optimization, decision support Database mining: converting (technical) records into knowledge Self-customizing programs: learning news filters, adaptive monitors Applications that are hard to program: automated driving, speech recognition Better Understanding of Human Cognition Cognitive science: theories of knowledge acquisition (e.g., through practice) Performance elements: reasoning (inference) and recommender systems Time is Right Recent progress in algorithms and theory Rapidly growing volume of online data from various sources Available computational power Growth and interest of AI-based industries (e.g., data mining/kdd, planning)
Relevant Disciplines Machine Learning Bayesian Methods Cognitive Science Computational Complexity Theory Control Theory Inference NLP / Learning Economics Neuroscience Bayes s Theorem Philosophy Missing Data Estimators Psychology Statistics Symbolic Representation Planning/Problem Solving Knowledge-Guided Learning PAC Formalism Mistake Bounds Planning, Design Optimization Meta-Learning Artificial Intelligence Game Theory Utility Theory Decision Models Bias/Variance Formalism Confidence Intervals Hypothesis Testing Power Law of Practice Heuristics Logical Foundations Consciousness ANN Models Learning
Application: Knowledge Discovery in Databases
Text Mining: Information Retrieval and Filtering 20 USENET Newsgroups comp.graphics misc.forsale soc.religion.christian sci.space comp.os.ms-windows.misc rec.autos talk.politics.guns sci.crypt comp.sys.ibm.pc.hardware rec.motorcycles talk.politics.mideast sci.electronics comp.sys.mac.hardware rec.sports.baseball talk.politics.misc sci.med comp.windows.x rec.sports.hockey talk.religion.misc alt.atheism Problem Definition [Joachims, 1996] Given: 1000 training documents (posts) from each group Return: classifier for new documents that identifies the group it belongs to Example: Recent Article from comp.graphics.algorithms Hi all I'm writing an adaptive marching cube algorithm, which must deal with cracks. I got the vertices of the cracks in a list (one list per crack). Does there exist an algorithm to triangulate a concave polygon? Or how can I bisect the polygon so, that I get a set of connected convex polygons. The cases of occuring polygons are these:... Performance of Newsweeder (Naïve Bayes): 89% Accuracy
Artificial Intelligence: Some Problems and Methodologies Problem Solving Classical search and planning Game-theoretic models Making Decisions under Uncertainty Uncertain reasoning, decision support, decision-theoretic planning Probabilistic and logical knowledge representations Pattern Classification and Analysis Pattern recognition and machine vision Connectionist models: artificial neural networks (ANNs), other graphical models Data Mining and Knowledge Discovery in Databases (KDD) Framework for optimization and machine learning Soft computing: evolutionary algorithms, ANNs, probabilistic reasoning Combining Symbolic and Numerical AI Role of knowledge and automated deduction Ramifications for cognitive science and computational sciences
A Generic Intelligent Agent Model Agent Sensors Internal Model (if any) Knowledge about World Knowledge about Actions Preferences Observations Predictions Expected Rewards Action Environment Effectors
Term Project Guidelines Due: 08 Dec 2004 Submit using new script (procedure to be announced on class web board) Writeup must be turned in on (for peer review) Team Projects Work in pairs (preferred) or individually Topic selection and proposal due 17 Sep 2004 Grading: 200 points (out of 1000) Proposal: 15 points Originality and significance: 25 points Completeness: 50 points Functionality (20 points) Quality of code (20 points) Documentation (10 points) Individual or team contribution: 50 points Writeup: 40 points Peer review: 20 points
Term Project Topics Intelligent Agents Game-playing: rogue-like (Nethack, Angband, etc.); reinforcement learning Multi-Agent Systems and simulations; robotic soccer (e.g., Teambots) Probabilistic Reasoning and Expert Systems Learning structure of graphical models (Bayesian networks) Application of Bayesian network inference Plan recognition, user modeling Medical diagnosis Decision networks or other utility models Probabilistic Reasoning and Expert Systems Constraint Satisfaction Problems (CSP) Soft Computing for Optimization Evolutionary computation, genetic programming, evolvable hardware Probabilistic and fuzzy approaches Game Theory
Homework 1: Machine Problem Due: 10 Sep 2004 Submit using new script (procedure to be announced on class web board) HW page: http://www.kddresearch.org/courses/fall-2004/cis730/homework Machine Problem: Uninformed (Blind) vs. Informed (Heuristic) Search Problem specification (see HW page for MP document) Description: load, search graph Algorithms: depth-first, breadth-first, branch-and-bound, A* search Extra credit: hill-climbing, beam search Languages: options Imperative programming language of your choice (C/C++, Java preferred) Functional PL or style (Haskell, Scheme, LISP, Standard ML) Logic program (Prolog) MP guidelines Work individually Generate standard output files and test against partial standard solution See also: state space, constraint satisfaction problems
Agent: Definition Any entity that perceives its environment through sensors and acts upon that environment through effectors Examples (class discussion): human, robotic, software agents Perception Signal from environment May exceed sensory capacity Sensors Acquires percepts Possible limitations Action Attempts to affect environment Usually exceeds effector capacity Effectors Transmits actions Possible limitations Intelligent Agents: Overview
How Agents Should Act Rational Agent: Definition Informal: does the right thing, given what it believes from what it perceives What is the right thing? First approximation: action that maximizes success of agent Limitations to this definition? Issues to be addressed now How to evaluate success When to evaluate success Issues to be addressed later in this course Uncertainty (in environment, in actions) How to express beliefs, knowledge Why Study Rationality? Recall: aspects of intelligent behavior (last lecture) Engineering objectives: optimization, problem solving, decision support Scientific objectives: modeling correct inference, learning, planning Rational cognition: formulating plausible beliefs, conclusions Rational action: doing the right thing given beliefs
Rational Agents Doing the Right Thing Committing actions Limited to set of effectors In context of what agent knows Specification (cf. software specification) Preconditions, postconditions of operators Caveat: not always perfectly known (for given environment) Agent may also have limited knowledge of specification Agent Capabilities: Requirements Choice: select actions (and carry them out) Knowledge: represent knowledge about environment Perception: capability to sense environment Criterion: performance measure to define degree of success Possible Additional Capabilities Memory (internal model of state of the world) Knowledge about effectors, reasoning process (reflexive reasoning)
Measuring Performance Performance Measure: How to Determine Degree of Sucesss Definition: criteria that determine how successful agent is Clearly, varies over Agents Environments Possible measures? Subjective (agent may not have capability to give accurate answer!) Objective: outside observation Example: web crawling agent Number of hits Number of relevant hits Ratio of relevant hits to pages explored, resources expended Caveat: you get what you ask for (issues: redundancy, etc.) When to Evaluate Success Depends on objectives (short-term efficiency, consistency, etc.) Is task episodic? Are there milestones? Reinforcements? (e.g., games)
Knowledge in Agents Rationality versus Omniscience Nota Bene (NB): not the same Distinction Omniscience: knowing actual outcome of all actions Rationality: knowing plausible outcome of all actions Example: is crossing the street to greet a friend too risky? Key question in AI What is a plausible outcome? Especially important in knowledge-based expert systems Of practical important in planning, machine learning Related questions How can an agent make rational decisions given beliefs about outcomes of actions? Specifically, what does it mean (algorithmically) to choose the best? Limitations of Rationality Based only on what agent can perceive and do Based on what is likely to be right, not what turns out to be right
What Is Rational? Criteria Determines what is rational at any given time Varies with agent, environment, situation Performance Measure Specified by outside observer or evaluator Applied (consistently) to (one or more) IAs in given environment Percept Sequence Definition: entire history of percepts gathered by agent NB: may or may not be retained fully by agent (issue: state and memory) Agent Knowledge Of environment required Of self (reflexive reasoning) Feasible Action What can be performed What agent believes it can attempt?
Ideal Rationality Ideal Rational Agent Given: any possible percept sequence Do: ideal rational behavior Whatever action is expected to maximize performance measure NB: expectation informal sense (for now); mathematical foundation soon Basis for action Evidence provided by percept sequence Built-in knowledge possessed by the agent Ideal Mapping from Percepts to Actions Figure 2.2, R&N Mapping p: percept sequence action Representing p as list of pairs: infinite (unless explicitly bounded) Using p: specifies ideal mapping from percepts to actions (i.e., ideal agent) Finding explicit p: in principle, could use trial and error Other (implicit) representations may be easier to acquire!
Structure of Intelligent Agents Agent Behavior Given: sequence of percepts Return: IA s actions Simulator: description of results of actions Real-world system: committed action Agent Programs Functions that implement p Assumed to run in computing environment (architecture) Hardware architecture: computer organization Software architecture: programming languages, operating systems Agent = architecture + program This course (CIS730): primarily concerned with p CIS540, 740, 748: concerned with architecture See also: Chapter 24 (Vision), 25 (Robotics), R&N Discussion: Real versus Artificial Environments
Agent Programs Software Agents Also known as (aka) software robots, softbots Typically exist in very detailed, unlimited domains Example (Real-time) critiquing, automation of avionics, shipboard damage control Indexing (spider), information retrieval (IR; e.g., web crawlers) agents Plan recognition systems (computer security, fraud detection monitors) See: Bradshaw (Software Agents) Focus of This Course: Building IAs Generic skeleton agent: Figure 2.4, R&N function SkeletonAgent (percept) returns action static: memory, agent s memory of the world memory Update-Memory (memory, percept) action Choose-Best-Action (memory) memory Update-Memory (memory, action) return action
Example: Automated Taxi Driver Agent Type: Taxi Driver Percepts Visual: cameras Profilometer: speedometer, tachometer, odometer Other: GPS, sonar, interactive (microphone) Actions Steer, accelerate, brake Talk to passenger Goals Safe, legal, fast, comfortable Maximize profits Environment Roads Other traffic, pedestrians Customers Discussion: Performance Requirements for Open Ended Task
Review: Course Topics Overview: Intelligent Systems and Applications Artificial Intelligence (AI) Software Development Topics Knowledge representation Logical Probabilistic Search Problem solving by (heuristic) state space search Game tree search Planning: classical, universal Machine learning Models (decision trees, version spaces, ANNs, genetic programming) Applications: pattern recognition, planning, data mining and decision support Topics in applied AI Computer vision fundamentals Natural language processing (NLP) and language learning survey Implementation Practicum 1-2 Students per Team
Terminology Artificial Intelligence (AI) Operational definition: study / development of systems capable of thought processes (reasoning, learning, problem solving) Constructive definition: expressed in artifacts (design and implementation) Intelligent Agents Topics and Methodologies Knowledge representation Logical Uncertain (probabilistic) Other (rule-based, fuzzy, neural, genetic) Search Machine learning Planning Applications Problem solving, optimization, scheduling, design Decision support, data mining Natural language processing, conversational and information retrieval agents Pattern recognition and robot vision
Summary Points Artificial Intelligence: Conceptual Definitions and Dichotomies Human cognitive modelling vs. rational inference Cognition (thought processes) vs. behavior (performance) Intelligent agent framework Roles of Knowledge Representation, Search, Learning, Inference in AI Necessity of KR, problem solving capabilities in intelligent agents Ability to reason, learn Applications and Automation Case Studies Search: game-playing systems, problem solvers Planning, design, scheduling systems Control and optimization systems Machine learning: pattern recognition, data mining (business decision support) Course Group: http://groups.yahoo.com/group/ksu-cis730-fall2004 More Resources Online Home page for AIMA (R&N) textbook: http://aima.cs.berekeley.edu CMU AI repository Comp.ai newsgroup (now moderated): http://groups.google.com