CS 360: Advanced Artificial Intelligence Lecture 1

Gautam Biswas Fall 2014

1 8/22/2009

Topics

Class overview Review chapter 1 (self-study) Review chapter 2 (self-study): discuss on 09/09

8/22/2009 2

Class/Instructor Information

Class time: Mon.–Wed. 8:45 am – 10:00 am Location: Room 129, Featheringill Hall (FGH) Instructor: Gautam Biswas Office: Room 366, Jacobs Hall (also in ISIS, 1025 16th Ave S, Room 401D) Tele: 343-6204 E-mail: gautam.biswas@vanderbilt.edu Office Hours: Mon-Wed.: 10:00 – 11:00 am

TA: Ye Cheng

8/22/2009 3

My Research Areas

Design and evaluation of Intelligent Learning Environments Combines AI, Cognitive Science, & Education

Machine Learning and Data Mining Anomaly Detection from Aircraft Flight Data (Big Data) Educational Data Mining

Multi-level distributed decision support systems Planning, scheduling, and resource allocation in distributed real time

environments; Uncertain decision making

Model-based reasoning for complex, dynamic systems Symbolic reasoning for uncertain dynamic processes Reasoning with Dynamic Bayes nets Applications to diagnosis and fault-adaptive control

8/22/2009 4

Text Book and References

Text: Stuart Russell and Peter Norvig, “Artificial Intelligence: A Modern Approach,” (AIMA), Prentice Hall, 3nd edition, 2010.

(http://aima.cs.berkeley.edu)

References: N.J. Nilsson, “Artificial Intelligence: A New Synthesis,”

Morgan Kaufmann, 1998. G.F. Luger, “Artificial Intelligence: Structures and Strategies

for Complex Problem Solving,” Addison Weseley, 2005. (newer edition ..)

8/22/2009 5

Programming Language

Assignments C, C++ Java Python etc.

8/22/2009 6

Grading Scheme

Home Work (4): 30% Programming Assignments (3): 30% Two Exams (Midterm + Final): 30%

(both take home) Project (Paper on chosen topic): 10%

8/22/2009 7

Reference Material

Additional reading material: papers – required reading Lecture slides & notes + HW + Programming Assignments + additional material

All material will be posted on Oak (course CS 360: Advanced AI)

8/22/2009 8

Home Work and Programming Assignments

Posted on Oak Must be submitted by due date, penalties for late submission Full, 75%, 50%, 0 credit rule

Has to be your own work Can discuss problem sets to understand nature of the problem No discussion of solutions, no sharing code Do not download solutions or program code from web, and submit your

own

Vanderbilt Honor Code http://www.vanderbilt.edu/student_handbook/Honor_System.htm Must put honor pledge on submitted assignments

8/22/2009 9

Chapter 1: What is AI?

What is AI? How do we go about doing AI? What does an AI researcher produce? What basic topics in AI are we going to cover in this course?

8/22/2009 10

What is Artificial Intelligence? Association for the Advancement of Artificial Intelligence (AAAI): “the scientific understanding of the mechanisms underlying thought and intelligent behavior and their embodiment in machines.” (www.aaai.org) John McCarthy: “It is the science and engineering of making intelligent machines, especially intelligent computer programs. It is related to the similar task of using computers to understand human intelligence, but AI does not have to confine itself to methods that are biologically observable.” Herb Simon: “AI can have two purposes. One is to use the power of computers to augment human thinking, just as we use motors to augment human or horse power. Robotics and expert systems are major branches of that. The other is to use a computer's artificial intelligence to understand how humans think. In a humanoid way. If you test your programs not merely by what they can accomplish, but how they accomplish it, they you're really doing cognitive science; you're using AI to understand the human mind."

8/22/2009 11

AI

Strong two way connection between AI ⇔ Cognitive Science

AI: computer-based models – knowledge representation (symbolic, mixed numeric and symbolic) and algorithms for problem solving Cognitive Science: models of human reasoning system

8/22/2009 12

What is Artificial Intelligence?

Study and learn about intelligent entities (understanding) Learn how to build them (building) Two aspects: develop a theory or systematic descriptions of intelligent

reasoning (science – conceptualize, theorize) write programs that make computers perform tasks

intelligently (engineering – build, verify, demonstrate

8/22/2009 13

What is Artificial Intelligence?

Two AI camps: 1. Strong AI 2. Weak AI

8/22/2009 14

What is Artificial Intelligence?

Claims of strong AI camp: computers can be made to think on a level (at least) equal to that of humans

What will this require? The Turing Test (Alan Turing, 1950)

Weak AI camp: add some thinking-like features to computers or computer programs

e.g., speech understanding systems, softbots and personalized assistants, experts systems that can analyze images or troubleshoot computers, drive by wire cars, autopilot of aircraft

8/22/2009 15

Additional Reading Computer Chess: Deep Blue Kasparov versus Deep Blue: http://www.cs.vu.nl/~aske/db.html http://en.wikipedia.org/wiki/Deep_Blue_(chess_computer) Deep Blue - Brute force computation or what?

http://www.byte.com/art/9707/sec6/art6.htm http://whyfiles.org/040chess/main2.html Games played by Deep Blue:

http://www.chessgames.com/perl/chessplayer?pid=29912

Robotics, Autonomous Control, and Planning DARPA 2005 Grand Challenge: http://www.darpa.mil/grandchallenge05/gcorg/index.html DARPA 2007 Grand Challenge:

http://www.darpa.mil/grandchallenge/overview.asp http://www.darpa.mil/grandchallenge05/TechPapers/Stanford.pdf

8/22/2009 16

Additional Popular Readings

Deep Learning (NY Times Article): posted on Oak

8/22/2009 17

Chapter 2: Intelligent Agents

Agents operate in an environment Sense + Act

Rationality Nature of environment Different types of Agents

8/22/2009 18

What is an intelligent agent? (Ch. 2)

Russell: anything that can perceive its environment through sensors and act upon its environment through effectors (actuators)

Environment

Agent

percepts sensors

actuators Actions

8/22/2009 19 Key idea: agent + environment

What is an intelligent agent?

Cognitive principles (think and act like a human) principles provided by Cognitive Science Some theory, mostly empirical

Principles of Rationality Logic + performance measure

Performance measure needs to be objective – therefore, defined as desired states of the environment as opposed to how the agent defines success

8/22/2009

20

Rational Behavior Depends on Performance measure that defines the criterion of success Agent’s prior knowledge of environment Actions that agent can perform Agents percept sequence to date

Autonomous Agent Less reliance on prior (pre-programmed) knowledge Improve performance with time

Rational Behavior ≠ Perfect Behavior

8/22/2009 21

Intelligent Agent

How do we define agent behavior? Percept sequence maps it into action(s) Agent function: map a percept sequence to action

Example, a table of percepts and corresponding actions Implementation of agent function as an agent

program Example, let’s build an agent that lives in a 2D

world, can navigate around obstacles Questions: how do you represent world (environment)? What are the percepts or percept sequences? What actions can the agent perform?

8/22/2009 22

Example Problem

1 1 1 1

1 1 1

1 1 1 1

1 1

1 1

Robot Agent

1 Objects (obstacles) Room Boundary

s1 s2 s3

s8 s4

s7 s6 s5

Robot can sense the 8 cells surrounding it, which are denoted by the 8 binary-valued variables, s1 to s8. For example, in the position on left, the sensed vector is

(0,1,0,0,0,1,0,0)

8/22/2009 23

Agent Behavior

Ideal Mapping from Percept Sequences to Actions Given set of all percept sequences and the set of actions the agent can perform, for each percept sequence determine the appropriate action the agent should perform. Results in an ideal mapping that is a table. Size of table = number of possible percepts Therefore, table can presumably be infinite Is this a good way to build an agent program?

8/22/2009 24

25

Task Environment Ideal Rational Agent: for each possible percept sequence, an ideal agent should do whatever is necessary to maximize its performance (based on some chosen measure), on the basis of the precept and whatever built-in knowledge the agent has.

Environment

Agent

percepts sensors

actuators Actions

Performance Measure Measure of success Has to be objective Not one measure works for all agents in all situations Can be multidimensional

Task Environment: P(erformance) E(nvironment) A(ctuators) S(ensors)

8/22/2009

Environments – define problem

Task Environment: P(erfromance) E(nvironment) A(ctuators)S(ensors) Agents perform actions on the environment This may/may not change environment, which then provides new percepts to the agent Environments can have different characteristics. Fully observable versus partially observable Deterministic versus stochastic Episodic versus sequential Static versus dynamic Discrete versus continuous Single versus Multi-agent

8/22/2009 26

27

Structure of Agent

Architecture + Program Architecture: sensors + actuators + computing device Program: implements the mapping from percept

sequences to actions Function of architecture – computing device does not have

much memory, typically program cannot deal with very long percept sequences

program generates walk action agent must have hardware for walking (e.g., legs)

8/22/2009

28

Agent Programs

Percepts determine actions. At each step, new percept sequence processed and action chosen. Performance measure external to agent. Agent types:

1) Reflex agents – actions based on current percept. Simple situations implemented as a table. Better implementation condition-action rules (example problem)

2) Model-based Reflex Agents – agents that keep track of the world (create maps, maintain internal states) old-state + new percept new state Action function (state,productions)

3) Goal-based agents: involve decision-making – which action will bring me closer to a desired goal: notions of search and planning

4) Utility-based agents: go beyond goals. If there are many ways to achieve goal, which one is best? Brings in notion of performance.

Performance may involve trade-offs. Quantized by means of utility function.

8/22/2009

Reflex Agent

8/22/2009 29

Environment

State of the world

Choose Action

Sensors

Actuators

Agent

Rules

Model-based Reflex Agent?

30

Agent Programs

Agent types: 5) Learning Agents: Idea – even if agent starts in unknown

environment (situation) it learns by observing, doing, etc. so that it becomes more competent as time progresses.

Environment

Critic

Learning Element

Problem Generator

Performance Element

Sensors

Actuators

feedback

learning goals

performance standard

changes

knowledge

8/22/2009

Agent

Reading Assignments for next class

Read Computer Chess & DARPA Urban Challenge articles Discussion next lecture Why are these problems/challenges of interest

to AI? Read about Deep Learning Discussion – what is new about Deep

Learning?

8/22/2009 31

Assignments for Next Class

What are – BFS, DFS, Iterative Deeping DFS, and Bidirectional Search? Similarities and Differences

8/22/2009 32

Next Class

Background reading: Chapter 3.1-3.4, Russell and Norvig Uninformed search: Depth and Breadth first search State space representation Comparisons

Topic for Next class: Informed (Heuristic Search Strategies) – Chapters 3.5-3.6, Russell and Norvig + additional papers A* (book) IDA* (book + Korf paper) Additional topics − Anytime Search (Hansen & Zhou paper)

8/22/2009 33