CS计算机代考程序代写 prolog deep learning flex AI algorithm 04Agents

04Agents

Agents and Introduction
to AI
CITS3001 Algorithms, Agents and Artificial Intelligence

2021, Semester 2Tim French
Department of Computer Science and Software Engineering
The University of Western Australia

Introduction

• We will consider what is meant by the terms
– Artificial intelligence
– Agents

• We will define
– Four ways of looking at the former
– Four general models for the latter

2

What is Artificial Intelligence
• Given that experts can’t even agree on a definition for the word “intelligence”, what does “AI” mean!?
• Movies

– Kubrick’s 2001: a Space Odyssey, 1968
(“I’m sorry Dave, I’m afraid I can’t do that”)

– Cameron’s The Terminator, 1984 (Skynet)
– Proyas’s I, Robot, 2004 (could you kill a robot?) http://www.bbc.com/future/story/20131127-

would-you-murder-a-robot)
• TV “science shows”

– Towards 2000, Beyond 2000, Beyond Tomorrow, …
• News/current affairs

– Deep Blue vs. Kasparov
– Watson vs. the best of the best
– AlphaGo vs Lee Sedol
– Japan: Robot to take top university exam http://www.bbc.com/news/blogs-news-from-elsewhere-

26418431
– Robots will be smarter than us all by 2029 http://www.independent.co.uk/life-style/gadgets-and-

tech/news/robots-will-be-smarter-than-us-all-by-2029-warns-ai-expert-ray-kurzweil-9147506.html
• Adverts

– Intelligent TVs, washers, cars, molecules…
3

AI in research….

4

Thinking Humanly

• Determine how humans think, and attempt to replicate it in software/hardware
• Develop a theory of the human mind, by one or more of

– Introspection
– Psychological experiments (top-down?)
– Brain imaging (bottom-up?)

• What level of abstraction is best?
– “Knowledge” or “circuits”?
– Should we model the “mind” or the “brain”?

• And how would we validate such a system?
• e.g. the General Problem Solver (GPS)

[Newell & Simon, 1961]
– Attempted to “solve like a human”
– No searching

• The question of whether machines can think … is about as relevant as the
question of whether submarines can swim Edsger Dijkstra

5

Acting Humanly

• Intelligence = the ability to act indistinguishably from a human in cognitive
tasks(?)

• An operational test: the Turing Test [Alan Turing 1950]

6

• H interrogates X in a black box
• If X is a computer, but H cannot tell,
X must be intelligent!

• Loebner Prize
• http://www.loebner.net/Prizef/loebner-

prize.html
• Basically an online Turing Test

• Botprize
• http://botprize.org/
• Can computers play like people?

• GECCO “humies”
• http://www.sigevo.org/gecco-

2014/humies.html
• Prizes for human-competitive results

Thinking Rationally

• Codify “laws of thought” or “right-thinking”
– Irrefutable reasoning processes
– Independent of what humans do

• All men are mortal
• Socrates is a man
• Therefore Socrates is mortal

• Captured in rules of inference
– modus ponens: (P Ù (P → Q)) → Q
– modus tollens: (¬Q Ù (P → Q)) → ¬P
– absorption: (P → Q) → (P → (P Ù Q))
– Many others

• Problems include
– Difficulty in codifying informal knowledge
– Difficulty in dealing with uncertainty
– Scalability issues

7

Acting Rationally

• Act in such a way as to achieve goals, given beliefs
• Define an agent, and give it

– Some goals
– The ability to perceive its surroundings
– The ability to perform actions
– The ability to “reason”

• It will (try to) find actions to achieve the goals
• Note that this doesn’t necessarily involve “thinking”

– e.g. is a thermostat intelligent?
• This gives us an engineering viewpoint

– Can we develop systems that do useful stuff?
– Or even cool stuff!?

• We have a proof of concept, after all
• Our view (the modern view) of AI is as the study, design, and construction of

intelligent agents
• For previous significant views, read up on the foundations and history of AI

– Section 1.2–1.4 of AIMA 8

So, What is an Agent?

• An agent
– Perceives its environment through sensors
– Acts on its environment through effectors

9

Rational Agents

• A rational agent tries to “do the right thing” wrt a set of goals or utilities
• The right thing can be specified by a performance measure defining a numerical

value for any environment history

• A rational action is whatever action maximises the expected value of the

performance measure, given the current state of the environment and the percept

sequence to date

• But note that

– Rational ≠ Omniscient

– Rational ≠ Clairvoyant

– Rational ≠ Successful

• It is entirely possible to do the right thing and to fail anyway

– Sometimes randomness is the most rational choice!

– e.g. games

• An agent’s behaviour is specified by an agent function mapping percept sequences
to actions

– The agent will usually also store knowledge or rules that help it to understand and to

select actions

• We will discuss four basic types of agents, in order of generality 10

Simple Reflex Agents

• Choose an action using condition-action rules
– e.g. if the car-in-front’s brake-lights come on

then apply your brakes
• The key word in this diagram is now

– No history is stored
– Some experts believe that this is how

simple life-forms (e.g. insects) behave

11

Model-based Reflex Agents

• While simply reacting to the (current) world is adequate in some circumstances,
most intelligent action requires more knowledge

– Stored memory of the past
– Understanding of the effects of actions

• Both of these require internal state
– e.g. you see a pedestrian ahead signal to a bus
– You know the bus will stop
– You should change lanes

• Also this allows much better for worlds that are only partially observable
– Which is by far the most common case

12

Goal-based Agents

• Reacting better to the changing world is
an improvement

– But what are we trying to achieve?
• Intelligent (and some other!) beings have
goals

– e.g. at a junction, which way do we turn?
• Achieving goals involves predicting the

future
– If I do this action, how will that change

the world?
• Some goals are simple

– Star Trek: boldly go where no one has
gone before

• Other goals are complex and require
planning

– Star Wars: defeat the Empire!
• Planning is fundamental and usually

requires search 13

Utility-based Agents

• A goal is a binary thing
– Achieve it or fail!

• Most outcomes are more continuously-measured
– e.g. which action will make me happier? Or richer?

• Usually defined as a utility to be maximised
– cf. optimisation problems

• Again partial observability rears its head
– The agent will try to maximise expected utility

14

Aspects of AI: Natural Language
Processing

• Natural language processing is the process of applying
meaning to text.

• More than just transcribing spoken word, this requires a
machine to understand the intent and meaning behind a
sentence, which requires semantic knowledge of the world.

• Closely related to machine translation, there has been
consistent effort in this area throughout the history of AI

• Rule based approaches, apply gramatical rules and pattern
matching approaches.

• Statistical approaches build statisticals models of meaning
by sampling large corpuses of text.

• Deep Neural Networks are the most recent and successful
approach and create embeddings of text into high
dimensional spaces where regions have semantic meaning.

Aspects of AI: Computer Vision

• Computer Vision has made exceptional advances in recent times, largely based on
the advent of convolutional nerual networks and deep learning.

• Images and videos are represented as high dimensional vectors, and passed
through many layers of a neural network.

• The are tagged (classified) and the neural network is trained to recognize the
classification via back propogation.

• Different layers of the neural network often correspond to recognizable features in
the image.

16

Aspects of AI: Optimization

• Optimization and Operations Research are branches of applied mathematics not
traditionally a part of artificial intelligence.

• However, many aspects of autonomous control and automation rely on a machine
selecting a good or best action.

• From an implementation point of view, this is often achieved by building abstract
simulations of the environoment, and trialling different actions to see which provides the
greatest utillity.

• A large part of this task is modelling and prediction: how should the simulated
environment respond to different actions, and what reward can we expect. Statistical
learning is often applied here.

• Once a model is provided, generic branch and bound/brute force methods can be used
to explore options.

Aspects of AI: Argumentation and
Reasoning

• Argumentation and reasoning is the process of applying
logical deductions and inferences to reach a conclusion from a
premise.

• It is an intrinsic part of bargaining, negotiation, and social
interactions.

• Despite the substantial advances in deep learning, recent
advances in reasoning have been much more modest.

• Reasoning is typically done via rule based systems, where
premises are transformed and match with patterns, where
deductions can be extracted.

• Programming systems like Prolog and LISP can encode such
rules, but then the reasoning typically proceeds through a
depth first search.

AI and job automation

• A lot of media focus has been on the opportunity of AI in the workforce, and the threat
to traditional vocations.

• Checkouts, bank tellers, typists are traditional professions that are already heavily
automated.

• There is a significant investment in autonomous road vehicles which could have a
massive impact on the workforce, but many legislative barriers remain.

• White collar automation refers to the process of automating routine tasks in accounting,
law, medicine and other traditional professional occupations. Predictive systems and
pattern matching can provide support to many of the typical tasks in these professions,
such as contract review, filing tax forms, or matching conditions to symptoms.