Comp Sci 110 - GR #3

AI

simulation of human intelligence by machines

What is the goal of AI?

to create systems that think and act like humans and rationally

Expert System

a computer that emulates the decision-making capabilities of a human expert

General AI

a form of AI that exhibits human cognitive abilities

General AI can think and accomplish tasks _______________

on their own

What are the capabilities of General AI?

logical thinking

plan

learn on its own

communicates in a natural language

Examples of General AI

every "cool" AI in movies (ex: terminator)

Narrow AI

a form of AI that is designed to perform a singular task or a limited set of tasks with high efficiency

Narrow AI are systems that can accomplish a wide range of problems but operate within?

a pre-defined range of functions

Capabilities of Narrow AI

cannot think for themselves

only perform tasks they were designed to accomplish

can learn when provided with training data

Examples of Narrow AI

Siri/Alexa

Telephone Menus

Video Game AI

Navigation Systems

Fully Observable AI

AI has access to all information

Partially Observable AI

AI can only see part of the picture

Competitive AI

AI competes against an agent with the opposite goal

Collaborative AI

AI can work together with agents

Discrete AI

AI with a finite number of possible states (decision trees)

Continuous AI

AI that continuously updates and improves their knowledge and capabilities based on new data without explicit retraining (Netflix's viewing recs)

Deterministic AI

outcome of an action is known with 100% certainty

Stochastic AI

the outcome of an action is known with some predictability

7 General AI Problem Types

Classification

Clustering

Optimization

Anomaly Detection

Ranking/Recommendations

Data Generation

Classification

using a set of training examples as a guide; categorize new inputs as belonging to one or more categories

Clustering

determine how to categorize or group data based on their common characteristics (EX: birthday photos, vacation photos)

How do Classification and Clustering differ?

Classification is using a training example (EX: husky) to categorize inputs (EX: husky vs wolf)

Clustering is grouping things together based on their similarities WITHOUT a training example (EX: photos making memories of you on vacation)

Optimization

given a function, determines what input(s) will result in the max/min value (EX: given your "recently watched" Netflix will suggest new movies that provide the most enjoyment)

Anomaly Detection

given a set of training examples, determines if an input is "out of the ordinary"

Ranking/Recommendations

given a series of inputs/previous actions, determines what will happen next (EX: Chess.com)

Data Generation

Produce novel data in order to achieve a specified goal (EX: inputting data into R-Script and having it produce a graph)

Sense

detects or is provided info about its environment

Think

considers possible actions and determines which one(s) it wants to perform

Act

performs the action chosen in the Think step

Which step is the most time-consuming for a Game AI? Why?

Thinking; it requires us to decide for the AI what actions to take

Hard-Coded Rules/Decision Trees

requires expert knowledge

Finite State Machine

a machine that consists of a fixed set of possible states and the transitions between them

Search Trees

depicts game states or a snapshot of the game at a given point of time and the actions that can be made at that time and the outcomes

Game State

snapshot of the game at a point

Breadth-First Search

look at all the nodes at one level (A, B, C)

Depth-First Search

pick a node and go as far down as possible (A --> A1, A2, A3...)

Heuristic

uses rules or methods to find solutions more quickly when other methods are too slow; "good enough"

Why do we use heuristics?

estimate the outcome of a decision without having to look multiple steps ahead

Algorithms

derive rules/behaviors from datasets

Datasets are used to:

train the computer

verify the computer learned the concept (not just the pattern)

Human Expertise

verify algorithms are working as intended

Machine Learning involves 3 things?

algorithms, datasets, human expertise

Model

a set of rules the AI learns from the data

Feature

columns in dataset; specific piece of information

Label

the concept we're training the model to predict; the result

Instance

row in dataset; complete set of features

Unsupervised Learning

algorithm is provided with an unlabeled dataset and is tasked with grouping similar instances

Supervised Learning

algorithm provided with a labeled dataset and is tasked with identifying the label by only looking at the features

Independent Variables

features

Dependent Variable

label

Supervised Learning Process

gather/prepare data

choose a model

train model

evaluate model

tuning

predicition

Gather and Prepare Data

what info we need, how we get that info, and how we organize it

Choose A Model

picking the right model requires understanding of the problem

Train Model

give training data to a model and let it learn

Evaluate Model

test model with data it has never seen before so we know it knows the actual concept

Confusion Matrix

where the model succeeds/fails

Tuning

adjust the settings of the AI

Epochs

how many times you show the training data to the AI and let it learn

Learning Rate

how much the algorithm changes when it makes a mistake

Prediction

use model in read world with real data

Narrow Focus

designed with a specific task; don't have general knowledge (EX: Deep Blue can beat anyone at chess but can't play checkers)

High Cost of Creation

requires humans to tell them what to do in every possible situation

Hidden Bias

AI's inherit bias from their creators (EX: Face ID favors white males)

Ambiguous Responsibility

if there is a fault in AI, we don't know if company or AI is responsible

Lack Interpretability

we don't know what a model has learned

Black Box

any system that cannot be directly observed and easily understood

Explainable AI

is artificial intelligence that is programmed to describe its purpose for making a decision

4 Ethical Criteria

transparent to inspection

predictability

robust against manipulation

responsibility

Network

a group of two or more computer systems linked together

Host

computer on a network

Client

computer trying to access the content on the server

Server

computer that stores content

Message

data we want to send

Protocols

how clients/servers talk to each other

HTTP (Hypertext Transfer Protocol)

use to request web pages

SMTP (Simple Mail Transfer Protocol)

clients use to SEND emails

FTP (File Transfer Protocol)

computers use to transfer files between each other

Who creates protocols?

IETF (Internet Engineering Task Force)

IP Address

The unique number assigned to each device on the Internet.

Public IP Address

identify a particular device on the public internet

Private IP Address

An IP address that cannot be routed over the Internet

Port

not physical; location on a computer where applications listen for incoming data

How many applications can listen on the same port?

1

TCP (Transmission Control Protocol)

ensures messages reach destination; packets can be out of order

UDP (User Datagram Protocol)

does not ensure the message makes it to the destination; packets stay in order

Circuit Switching

private dedicated channel is established for communication; constant reliable path; packets sent/received in order; no one can use it except sender and receiver

Cons of Circuit Switching

if link in circuit fails; all comms are disrupted

Packet Switching

packets travel from source to destination using best path; many comms can use the same path; if path fails, use another one;

Cons of Packet Switching

no guarantee of packets arriving; packets may be jumbled

Switch

connects multiple hosts/devices on the same network

Router

connects multiple networks; has many network addresses

IPv4

uses 8 bits for each of 4 parts

IPv6

uses 16 bits for each of 8 parts

Static

someone manually assigns specific IP addresses to a device

Dynamic

uses DHCP to assign IP addresses

NAT (Network Address Translation)

allows us to share a single IP with an entire private network

DNS (Domain Name System)

converts IP to network names (google.com)

WEP/WPA/WPA2

encrypts traffic, but anyone with key can see data

HTTPS

ensures info you send is private