Measuring Universal Intelligence By: Tyler Staudinger



Similar documents
What is Artificial Intelligence?

COMP 590: Artificial Intelligence

CSC384 Intro to Artificial Intelligence

CS440/ECE448: Artificial Intelligence. Course website:

CSE 517A MACHINE LEARNING INTRODUCTION

Artificial Intelligence (AI)

Who needs humans to run computers? Role of Big Data and Analytics in running Tomorrow s Computers illustrated with Today s Examples

CHAPTER 15: IS ARTIFICIAL INTELLIGENCE REAL?

COMP-424: Artificial intelligence. Lecture 1: Introduction to AI!

A Client-Server Interactive Tool for Integrated Artificial Intelligence Curriculum

BBC LEARNING ENGLISH 6 Minute English Do you fear Artificial Intelligence?

TD-Gammon, A Self-Teaching Backgammon Program, Achieves Master-Level Play

Watson. An analytical computing system that specializes in natural human language and provides specific answers to complex questions at rapid speeds

Appendices master s degree programme Artificial Intelligence

Turing Machines, Part I

Game Playing in the Real World. Next time: Knowledge Representation Reading: Chapter

Computer Science Theory. From the course description:

Introduction to Artificial Intelligence

Game playing. Chapter 6. Chapter 6 1

comp4620/8620: Advanced Topics in AI Foundations of Artificial Intelligence

Applications of Artificial Intelligence. Omark Phatak

Turing Tests with Turing Machines

About the Author. The Role of Artificial Intelligence in Software Engineering. Brief History of AI. Introduction 2/27/2013

Acting humanly: The Turing test. Artificial Intelligence. Thinking humanly: Cognitive Science. Outline. What is AI?

Beating the MLB Moneyline

EXECUTIVE SUPPORT SYSTEMS (ESS) STRATEGIC INFORMATION SYSTEM DESIGNED FOR UNSTRUCTURED DECISION MAKING THROUGH ADVANCED GRAPHICS AND COMMUNICATIONS *

A Robustness Simulation Method of Project Schedule based on the Monte Carlo Method

MAN VS. MACHINE. How IBM Built a Jeopardy! Champion x The Analytics Edge

Maximizing Precision of Hit Predictions in Baseball

21 st Century Knowledge Worker: the Centaur

CSCI Historical Development. May 29, 2015

Super Turing-Machines

Predictive Act-R (PACT-R)

Lab 11. Simulations. The Concept

Telling Humans and Computers Apart (Automatically) How Lazy Cryptographers do AI

Characteristics of Computational Intelligence (Quantitative Approach)

2015 State of Artificial Intelligence & Big Data in the Enterprise

MATH 140 Lab 4: Probability and the Standard Normal Distribution

Discrete Mathematics and Probability Theory Fall 2009 Satish Rao, David Tse Note 10

Draft dpt for MEng Electronics and Computer Science

Applying Deep Learning to Car Data Logging (CDL) and Driver Assessor (DA) October 22-Oct-15

Artificial Intelligence

Regular Languages and Finite Automata

Elementary Statistics and Inference. Elementary Statistics and Inference. 16 The Law of Averages (cont.) 22S:025 or 7P:025.

Learning What to Value

Roulette-Tools PATTERN RECOGNITION TRAINING

Complement. If A is an event, then the complement of A, written A c, means all the possible outcomes that are not in A.

NATIONAL: THE GOOD AND MOSTLY BAD OF ARTIFICIAL INTELLIGENCE

Appendices master s degree programme Human Machine Communication

How To Use Neural Networks In Data Mining

Machine Learning Introduction

The Peruvian coin flip Cryptographic protocols

Driving Insurance World through Science Murli D. Buluswar Chief Science Officer

6.080 / Great Ideas in Theoretical Computer Science Spring 2008

Why Sales Training Succeeds... Or. Fails. By Ron Willingham

Probability Using Dice

Master of Science in Computer Science Information Systems

Andre Standback. IT 103, Sec /21/12. IBM s Watson. GMU Honor Code on I am fully aware of the

Study Plan for the Master Degree In Industrial Engineering / Management. (Thesis Track)

What is the Probability of Pigging Out

2013 International Symposium on Green Manufacturing and Applications Honolulu, Hawaii

Likely Developments in ODR Notes Matthew Lavy and Richard Susskind July Introduction

Intelligent Agents. Based on An Introduction to MultiAgent Systems and slides by Michael Wooldridge

Reinforcement Learning and the Reward Engineering Principle

An Introduction to. Metrics. used during. Software Development

Section 6.2 Definition of Probability

Optimizing content delivery through machine learning. James Schneider Anton DeFrancesco

Introductory Probability. MATH 107: Finite Mathematics University of Louisville. March 5, 2014

How To Understand The Relation Between Simplicity And Probability In Computer Science

Introduction to Turing Machines

Basic Data Analysis. Stephen Turnbull Business Administration and Public Policy Lecture 12: June 22, Abstract. Review session.

Master's projects at ITMO University. Daniil Chivilikhin PhD ITMO University

The Relationship between Artificial Intelligence and Finance

INTRODUCTION TO MACHINE LEARNING 3RD EDITION

cs171 HW 1 - Solutions

Maths Targets for pupils in Year 2

Lesson Plans for (9 th Grade Main Lesson) Possibility & Probability (including Permutations and Combinations)

Artificial Intelligence for ICT Innovation

Reinforcement Learning

A Statistical Analysis of Popular Lottery Winning Strategies

UNDERGRADUATE DEGREE PROGRAMME IN COMPUTER SCIENCE ENGINEERING SCHOOL OF COMPUTER SCIENCE ENGINEERING, ALBACETE

Holland s GA Schema Theorem

- 1 - intelligence. showing the layout, and products moving around on the screen during simulation

Describe the process of parallelization as it relates to problem solving.

Gifted & Talented Program Description

Gibbs Sampling and Online Learning Introduction

PETITION/PROGRAM SHEET Degree: Bachelor of Science Major: Computer Science

Accelerated Bachelor of Science/Master of Science in Computer Science. Dual Degree Program

College of information technology Department of software

Graduate Co-op Students Information Manual. Department of Computer Science. Faculty of Science. University of Regina

CS Master Level Courses and Areas COURSE DESCRIPTIONS. CSCI 521 Real-Time Systems. CSCI 522 High Performance Computing

Probabilistic Strategies: Solutions

Course 395: Machine Learning

PhD in Computer Science at North Carolina A&T State University

Review on Using Packages to Enhance the Teaching of Critical Path Networks

MULTIPLE CHOICE FREE RESPONSE QUESTIONS

Risk Analysis Overview

15-381: Artificial Intelligence. Probabilistic Reasoning and Inference

Machine Learning: Overview

Hyper-connectivity and Artificial Intelligence

Transcription:

[1] Measuring Universal Intelligence By: Tyler Staudinger

Overview How Do we Define Intelligence? Current Tests of Machine Intelligence The Ideal Test for Intelligence Components of an Intelligence Test An Example Algorithmic Information Theory The Universal Intelligence Test The Anytime Universal Intelligence Test The Future of Artificial Intelligence

How Do We Define Intelligence? Intelligence is the ability of a rational agent to perceive its environment, and take actions that maximize its chances of success in that environment. A rational agent is an entity capable of logical decision making processes. Create intelligence, not evaluate it. [2]

Landmarks in Artificial Intelligence 1997 -The IBM supercomputer Deep Blue beat a reigning world chess champion, Gary Kasparov 2005- A Stanford robot won the DARPA Grand Challenge by driving autonomously for 131 miles along a desert trail 2011 The IBM question answering system Watson defeats two champions in the game Jeopardy! [8]

Current Tests of Machine Intelligence What is a Turing Test? Such a test has a number of shortcomings: It is anthropomorphic It does not give a score It is relatively easy to cheat It requires a human judge. [5] CAPTCHA ( Completely Automated Public Turing Test to tell Computers and Humans Apart) CAPTCHA s are easy questions for humans hard for A.I., i.e. Character Recognition In 10-20 years CAPTCHA s will be hard for humans

The Current State of Machine Intelligence There is no system which is able to adapt to many environments, this is commonly called a Strong A.I. All the progress in Machine Intelligence has only addressed specific problems Even when considering the abilities of Watson, there is no true understanding behind its actions, it is simply a highly advanced search engine There is also no true test of Intelligence that has been developed Until Now

The Ideal Test for Intelligence Universal (Biological or Artificial) Quantitative ( Based upon Information Theory and probability) Can evaluate any past, present, or future system Can evaluate any intelligence level (brilliant/inept), or time scale (fast/slow) Time relative, longer test gives better estimate [3]

Components of an Intelligence Test Three parameters to make an intelligence measurement A subject to be examined, this is known as an agent. A structure that the agent operates in, this is known as an Environment A protocol from which a performance score can be derived Turing Test and CATCHA address the first two but do not give a score This can be addressed by adding a reward when the agent performs the desired action

The Interaction Between Agent and Environment Observation The agent inputs an action to the environment The environment outputs: An observation, and a reward This is an example of reinforcement learning Agent Reward Action Environment

Example: The Monkey and The Banana Consider a simple game 3 Buttons each under a cell meant to hold a ball One cell has a ball in it If the monkey hits the button under the cell with a ball in it, they get a banana From the definitions we just developed the rewards in this environment are: Now to evaluate the monkeys performance Performance is the Expected Value of all the rewards:

Hypothetical Agent Behaviors Some agents are able to learn an environment quickly while others cannot Some receive more reward in the beginning but receive less reward overall, while the opposite may be true for other agents. [1]

Problems with this Model Not every environment is the same, need quantify the complexity of an environment Should use many environments of varying complexity Complex environments should be weighted more Rewards must be balanced to penalize random guessing (more on this later) Time needs to be taken into account [6]

Quantifying Environment Complexity Algorithmic Information Theory relates the notions of computation and complexity Kolgmogorov complexity is the length of the shortest program p that outputs a given string x over a machine U. [1] Lets look at an example: This idea can easily be applied to an environment, in an environment the complexity is the shortest sequence of actions that will generate a reward.

Incorporating Environment Complexity into The Intelligence Measure Now we can come up with a definition of intelligence, that incorporates multiple environments with varying complexity [1]

Penalizing Random Behavior To ensure that an agents behavior is not simply random a balanced environment can be used. This can be accomplished by including penalties as well as rewards This allows for the quick identification of agents with random behavior

Incorporating Time into the Formulation The time to complete the test is finite so this must be taken into account The more time that elapses the better the estimate given [1]

The Final Anytime Intelligence Test Algorithm [1]

The Potential Results of the Test for Various Agents This system has all of the characteristics we stated for an ideal test It can be used for humans or an agent more intelligent than humans [1]

Problems with the Test and the Future of Machine Intelligence The test developed is much more quantitative and rigorous than its predecessors Environments would have to be defined by humans, this can introduce a source of error. There aren t any agents yet, that are more intelligent than humans As of yet there is no machine intelligence even close to that of a human Machines intelligences are specialized at one task and are not very versatile in a variety of environments like humans are. Perhaps one day a machine intelligence will be surpass humans but that will not happen in the near future. It may be possible to pass the Turing Test but true intelligence is a long way off [7]

ECE Core Topics: Algorithms, Probability, and Discrete Math An Algorithm is a set of well defined steps to accomplish a task The Conditional Probability (A B) is the probability of observing A given that B has occurred For example if we have 2 dice and we roll a 5 on one of them the conditional probability of obtaining a 7 when the second die is rolled is 1/6. Discrete Math: the realm of mathematics that deals with non-continuous functions and variables

Questions? [4]

Works Cited [1] Measuring universal intelligence: Towards an anytime intelligence test,artificial Intelligence Volume 174, Issue 18, December 2010, Pages 1508-1539 [2]A.M. Turing, Computing machinery and intelligence, Mind 59 (1950), pp. 433 460. Full Text via CrossRef [3]M. Li and P. Vitányi, An Introduction to Kolmogorov Complexity and Its Applications (3rd ed.), Springer-Verlag, New York (2008). [4]D.L. Dowe, A.R. Hajek, A computational extension to the Turing test, in: Proceedings of the 4th Conference of the Australasian Cognitive Science Society, University of Newcastle, NSW, Australia, 1997. [5]D.L. Dowe and A.R. Hajek, A computational extension to the Turing test Technical report #97/322, Dept. Computer Science, Monash University, Melbourne, Australia, 9 pp. http://www.csse.monash.edu.au/publications/1997/tr-cs97-322-abs.html. Images: [1] http://www.hplusmagazine.com/articles/ai/how-long-till-human-level-ai [2] http://web.mit.edu/newsoffice/2009/ai-overview-1207.html [3] http://www.soumyabrata.com/article/intelligent-machines-artificial- Intelligence.aspx [4] http://www.huffingtonpost.com/dr-orin-levine/questions-as-milepostsof_b_448870.html [5] http://www.watblog.com/wp-content/uploads/2010/09/captcha.png [6] http://lucalongo.allrightsolutions.com/upload/images/distributed_artificial_intelligence.jpg [7] http://3ddigitalwallpapers.com/wp-content/uploads/2008/05/artificial-intelligence.jpg [8] http://www.nicholastam.ca/wp-content/uploads/2011/02/watson-ai-jeopardy.jpg