CS 510: Intro to Artificial Intelligence

Rachel GreenstadtDepartment of Computer Science

Drexel Universitywww.cs.drexel.edu/~greenie/cs510/

Overview

• What is Artificial Intelligence?

• History of AI

• What is CS 510?

• Syllabus, Schedule, Grading

• Final Project

• Overview of AI Topics

Introductions

• Introduce yourself:

• Your name

• Undergrad/Masters/Ph.D/How many years at Drexel?

• What is your research area?

• Which faculty member(s) do you work with?

• What brings you to CS 510?

• What else should we know about you? :)

What is AI?

Class Exercise

• Answer the following questions:

• What is Intelligence?

• What is Artificial Intelligence?

• What is an agent? What attributes does an agent have?

• When you’re done, swap your answers with a neighbor

A 42

Each card has a number or letter on one sideand a square or circle on the other side

Which cards must you turn over to determine if the following statement is true:

Every card with a letter on one side has a square on the other side

Thinking like humans

• 90% of humans get it wrong

• Answer is cards 2 and 3

• Most people pick 1 and 3

20 yrs Beer24 yrs Cola

Each card has an age on one sideand a drink on the other side

Which cards must you turn over to determine if the following statement is true:

Everyone in the bar is following the law.

What is AI?

Thinking like a human

Thinking rationally

Acting like a human

Acting rationally

Why Study AI?

• Fundamental scientific questions

• What does it mean to be smart?

• What makes us smart?

• Can our intelligence be replicated or exceeded? And how?

Why Study AI?

• Fundamentally useful engineering question

• AI in computers increases humanity’s collective intelligence and abilities

• Areas where computers lack the ability to act rationally limit us

Why Study AI?

But is it even possible?

• Billions of human computers must be doing something....

• Strong vs. Weak AI

• Human-level intelligent machines, conscious?

• “thinking-like” features to make computers more useful

agent

1. One that acts or has the power or authority to acts

2. One empowered to act for or represent another

15

Agents

Simple reflex agent

Modern AI Agents

• Not just AI, but AI situated in some environment

• Not just inference, but inference used in some context

• Not just a control loop, but complex autonomous decision-making

• Not just an algorithm, but an intelligent system

• Holistic approach to AI

• Multiple AI tools can be integrated to build an Agent

Intelligent Software Agents

• Responsive

• Goal-Directed

• Autonomous

• Social

19

PEAS

• Performance Measure

• Environment

• Actuators

• Sensors

• Examples?

Autonomous Cars

• Consider an automated taxi driver:

• Performance measure: Safe, fast, comfortable trip, maximize profits

• Environment: Roads, other traffic, pedestrians, customers

• Actuators: Steering wheel, accelerator, brake, signal, horn

• Sensors: Cameras, sonar, speedometer, GPS, odometer, engine sensors, keyboard, lidar

Agent or Program?

The easy stuff is hard

• Computers still can’t speak, see, or reason like a 5 year old child

• And the hard stuff is easy....

• Playing chess

• Proving theorems

• Diagnosing medical conditions

But huge advances in perceptual AI lately!

Task Hierarchy

• Perceptual and manipulation intelligence

• Emotional intelligence

• Social intelligence

• Cognitive / Reasoning tasks

25

AI Historical Highlights• 5th century

• Aristotle invents syllogistic logic

• 13th century

• zairja device used by Arab astrologers to calculate ideas mechanically

• Ramon Llull creates Ars Magna theological argumentation device

• 17th century

• Material arguments for thinking: Hobbes, Descartes

• Pascal invents mechanical calculating device

• 19th century

• Babbage and Lovelace work on programmable mechanical machine

• Boolean algebra representing some “laws of thought”

AI Historical Highlights• 1928 von Neuman’s minimax algorithm, used for game-playing

• 1950 Turing test devised

• 1950 Asimov publishes the 3 laws of robotics

• 1956 McCarthy coins “Artificial Intelligence” / Dartmouth conference

• Early years (1956-1970)

• Micro-worlds

• Reasoning by search

• Many successes, lots of optimism/hype - Samuel’s checkers, Gelemter’s Geometry theorem prover, Shakey, Dendral

AI Historical Highlights• AI Winter (1970s)

• Perceptrons - limits of neural networks

• language difficulties - “The spirit is willing but the flesh is weak” ==> “The vodka is good but the meat is rotten”

• Development of computational complexity

• Loss of funding

• AI becomes an industry (1980s)

• Expert, intelligent systems all the rage

• Bubble happens and expectations raised again

AI Historical Highlights• 2nd AI Winter (late 1980s - 1990s)

• More disappointment as AI fails to make people rich

• Expert systems are “brittle”

• Funding cut again

• AI Becomes a Science/Intelligent Agents (1987-present)

• Victory of the “neats” (vs “scruffies”)

• Statistical machine learning/HMMs has many successes

• AI starts to make people rich

• Moore’s law makes a lot more possible

• Emergence of Intelligent Agent approach

• Availability of very large data sets / deep learning

Now

• Boom cycle

• Disappointment to doomsday overnight?

• Where is the next bust/disappointment?

• Interplay between tacit and explicit knowledge?

• Working with humans?

• Safety/ethics/accountability

AI State of the Art

AI Applications

AI in Space

Autonomous satellite separation and docking

Exploring MarsMonitoring the sky

with telescope arrays

AI Art

What is CS 510?

Course Information

• Textbook

• Stuart Russell and Peter Norvig

• Artificial Intelligence: A Modern Approach

• Prentice-Hall (Third Edition)

• Supplementary Readings

• Available on course website

• http://www.cs.drexel.edu/~greenie/cs510

Course Objectives

• Learn about AI techniques

• Learn how to do AI research (grad class)

Expectations and Policies

• Use CS Department academic integrity policy – linked from course website

• When in doubt, be transparent, list collaborators and sources, ask if it’s ok

• Exams will cover materials from lectures and readings

• Two late days for assignments, all other late material 20% off per day (for group assignments, late days are for all group members). If you use up your late days early, don’t expect extensions.

Schedule

• Intro to AI

• Search and Problem Solving

• Planning

• Knowledge Representation

• Learning

Evaluation

• 20% Online Exams

• Midterm 10%

• Final 10%

• 20% Programming Assignments

• 20% Class Participation

• 40% Final Project

Homeworks

• HW 1

• Search with sliding block puzzles

• C/C++ or python

• HW 2

• Sentiment analysis with bayesian learning

• Python

Sliding Block Puzzle

• Implement program to solve sliding block puzzle using search

• Start early, do the setup functions

• Extra credit: heuristic search

Class Participation

• In-class exercises

• Class discussions

• BBLearn Online discussions

• More instructions on website

Readings and Discussion

• Expectations for all students

• Read papers before class, come ready to discuss

• Send two discussion points/questions to course board by *Friday* before class

• Discussion points should be twitter short (140 chars)

• Online students

• By day of class 3 pm

• Reply to two discussion points before class (1-3 sentences)

Final Project Read Handout!

• Free form research project

• Groups of 2-3 people

• Topic related to AI

• Milestones

• Groups and topic (Oct 8)

• Proposal due (Oct 22)

• Presentation (Dec 3)

• Project write up (Dec 3)

The Final Project Proposal

• 2 pages long

• Problem Statement and Motivation

• Brief Description of Approach

• Related Work and novelty

• Evaluation approach

• Milestones

AI Topics

Topics

• Week 2 (10/1)

• The AI Enterprise (Turing)

• Search RN Ch 3, 4

• Week 3 (10/8)

• No class

• Project pre-proposal due

• Week 4 (10/15)

• Multiagent Systems (Sycara)

• Google (Page, et al.)

• Contraints (RN Ch 6)

• Homework 1 due

• Week 5 (10/22)

• Games and game theory (RN Ch 5, 17.5-17.6)

• AlphaGo (Silver et al.)

• Project Proposal due

• Week 6 (10/29)Online Midterm

• Week 7 (11/5)

• BDI (Tambe)

• Intelligence without representation (Brooks)

• Logic RN Ch 7, 8

• Week 8 (11/12)

• Machine learning RN Ch 18

• Adversarial classification (Dalvi et al.)

• Week 9 (11/22)

• Planning (RN Ch 10, 11)

• Week 10 (11/29)

• Homework 2 due

• Bias (Caliskan et al.)

• Week 11 (12/3)

• Project presentations

• Project due

Search• The “Heuristic Search Hypothesis”

- (Newell and Simon)

• Subroutine of intelligent systems

• problem solving

• planning

• knowledge

• games

Some search issues we’ll discuss

• Intractability of exhaustive search

• Use of heuristics (A*)

• Local search “satisficing”

Open Problems

• Distributed search

• Dynamic search

• 2010 class: A Travel-Time Optimizing Edge Weighting Scheme for Dynamic Re-planning.  Andrew Feit, Lenrik Toval, Raffi Hovagimian and Rachel Greenstadt. AAAI 2010 Workshop on Bridging The Gap Between Task And Motion Planning (BTAMP)

• http://www.seas.upenn.edu/~maximl/wt/AAAI10_ws/BTAMP10_schedule.html

Constraint Reasoning

• Way of representing knowledge and structure on a problem so that standard heuristics can be applied

• Problems expressed as:

• Set of variables that need values

• Set of domains from which the values are drawn

• Set of constraints that represent relationships between the variables (must be satisfied or optimized)

Applications

• Supply chain management

• Scheduling

• Resource and task allocation

• Multiagent coordination

Open problems in Constraint Reasoning

• How to easily express problems as constraint problems

• What if the domain is dynamic or uncertain?

• How do you measure performance in distributed systems?

• See the Constraint Programming (CP) conference or the Distributed Constraint Reasoning (DCR) workshop

Games / Adversarial Search

• Inherently multiagent and competitive

• Classic work in turn based games

• Chess

• Checkers

• Go

• Now poker, general game playing

• http://www.computerpokercompetition.org/

• http://games.stanford.edu/

Mechanism Design

• Construct incentives for agents that are:

• self-interested

• utility-maximizing

• Applications

• Auctions

• Reputation systems

• Traffic systems

Knowledge Representation

• What is common sense?

• How a problem is represented greatly affects its efficiency

• How can we encode the things we know so computers understand them?

• How can representations be biased?

• Word embeddings

• https://www.tensorflow.org/tutorials/word2vec

• conceptnet.io

Model-based Reflex Agent

Planning

• Given

• a set of actions

• a goal state

• a present state

• Choose actions to get to the goal state

• And what if you have a team of agents...

Goal-based Agent

Utility-based Agent

Planning problems

• Planning in the real world

• Highly dynamic environments

• Uncertain information

• How can plans be recognized?

• Games (Poker? Football?)

Learning

• What does it mean for computers to learn?

• Supervised

• Unsupervised

“circle” “square” “circle” “square” …

“group these into two categories”

Learning Agent

• Predicting community ratings on web forums and blogs

• Authorship recognition

• learning who wrote a document by linguistic style

• Experiment with applying to text messages/transcribed speech

Learning Projects

Adversarial Learning

• cleverhans.io

•

Project Ideas

• Can we use stylometry to assess machine translation quality?

• Could you use adversarial examples to build an instagram/Facebook filter that isn’t machine-readable?

• Could something like Katia Sycara’s object salience maps be useful to help detect/prevent adversarial images?

• Can we use Caliskan’s methods to measure the level and type of bias in various textual sources (news sources?)

• Could we fine-tune word embeddings to remove bias?

• Can we build a theory-of-mind (based on beliefs, desires, and intentions?) of the user to help make security decisions?

Resources

• www.aitopics.org

• http://aima.cs.berkeley.edu

• http://library.drexel.edu

• http://aispace.org

Readings for next week:

• Turing, A.M. (1950). Computing machinery and intelligence. Mind, 59, 433-460.

• Initial discussion comments due Friday at 6 pm

• Online replies due Monday at 3 pm.