AI Week 1

Question 1

Definition of AI

Answer 1

AI is when a computer acts rationally

Question 2

Agent VS Rational Agent

Answer 2

Agent: anything that perceives its environment through sensors and acts upon that environment through actuators.

Rational Agent: an agent that achieves its goal while optimizing some performance measure

Question 3

Dartmouth Meeting

Answer 3

A meeting of 10 researchers at Dartmouth in 1956 to discuss AI. Considered the birthplace of modern AI.

Question 4

How does an agent interact with its environment?

Answer 4

Percept/Sensors: The agent can perceive or sense the environment

Actuators/Actions: The agent can execute an action in the environment

Question 5

Types of Environments

Answer 5

Remember with the mnemonic “D-SOAKED”:

Deterministicness (deterministic or stochastic):
Staticness (static or dynamic)
Observability (fully observable or partially observable or unobservable)
Agency (single agent or multi-agent)
Knowledge (known or unknown)
Episodicness (episodic or sequential)
Discreteness (discrete or continuous)

Question 6

Deterministic & Stochastic Environment

Answer 6

An environment is deterministic if the next state is perfectly predictable given knowledge of the previous state and the agent’s action. Ex: chess AI.

An environment is stochastic if we can only predict the probablity of the next state occuring given knowledge of the previous state and the agent’s action. Ex: we know the probablity a self-driving car will crash into someone.

Mathematically:
Deterministic: transition_model(state1, action) = state2
Stochastic: transition_model(state1, action) is a probability measure

Question 7

Static & Dynamic Environment

Answer 7

An environment is static if it does not change while the agent deliberates. Ex: chess game.

An environment is dynamic if it does change while the agent deliberates. Ex: other cars move while our self-driving car decides to brake.

Question 8

Fully Observable, Partially Observable, & Unobservable Environment

Answer 8

An environment is fully observable if the agent can perceive and access all the information relevant to executing an action. Ex: chess AI. Math: the set S of possible percepts contains all percepts relevant to selecting an action

An environment is partially observable if the agent can perceive and access some of the information relevant to executing an action. Ex: self-driving cars. Math: the set S contains some but not all the percepts relevant to selecting an action

An environment is unobservable if the agent can perceive and access none of the information relevant to executing an action. Ex: no examples of this in the real world. The set S contains no percepts and is an empty set.

Question 9

Single-Agent & Multi-Agent Environment

Answer 9

An environment is a single agent environment if there is only one decision-making entity in the environment. Ex: image classification because there is only one decision-making entity.

An environment is a multi-agent environment if there are multiple decision-making entities in the environment. Ex: a self-driving car is one decision-making entity and a pedestrian is another such entity.

Question 10

Known & Unknown Environment

Answer 10

An environment is a known environment if the results for all actions are known to the agent. Ex: the agent knows all possible results of moving a chess piece in a chess game.

An environment is an unknown environment if the results for all actions are not known to the agent. Ex: a self-driving car does not know all possible results of coming to a shortstop as the driver behind us may react.

Math:
S1 = set of all possible initial states
A = set of all possible actions
S2 = set of all possible resulting states
transition_model: S1 x A -> S2
known: the set S2 contains all possible relevant resulting states
unknown: the set S2 does NOT contain every possible resulting state

Question 11

Episodic & Sequential Environment

Answer 11

An environment is an episodic environment if actions do not require knowledge of the past and rely only on the current percept. IOW, all actions are executed as stand-alone episodes. Ex: an image classification AI only relies on the current image it senses to classify it.

An environment is a sequential environment if actions do require knowledge of the past and rely on the percept sequence. Ex: self-driving car relies upon the knowledge of both where the pedestrian was before and where it is now to determine the speed of the pedestrian and thus if the car needs to brake to avoid hitting him

Question 12

Discrete & Continuous Environment

Answer 12

An environment is a discrete environment if all of its precepts are discrete. Ex: The (x,y) position of a chess piece perceived with a camera.

An environment is a continuous environment if all of its precepts are continuous. Ex: The velocity of a self-driving car.

Question 13

Agent Function

Answer 13

A function that maps a percept sequence to an action

Question 14

Agent Program

Answer 14

While an agent function is an abstract mathematical function, the agent program is a concrete implementation of the agent function.

An agent program takes in the current percept sequence and outputs an action.

Question 15

Simple Reflex Agent

Answer 15

An agent that selects an action on the basis of the current percept, ignoring the rest of the percept history.

Agent functions for simple reflex agents are usually written as if-then statements.

Question 16

Model-based Reflex Agent

Answer 16

The simple reflex agent only makes decisions based on the current percept sequence. But the model-based reflex agent makes decisions based on the entire percept sequence. To store and understand this percept sequence, the model-based reflex agent maintains some sort of ‘internal state’ which reflects part of the unobserved environment. This agent still uses primarily if-then statements.

Agents maintain an ‘internal state’ with

1) transition model, which captures the effects of the agent’s actions on the world. maps actions and the current state to a future state.
2) sensor model, which captures how the percepts represent what is occurring in the world. maps percepts to the current state of the world. Ex: maps a droplet-shaped object that the camera detects to rain.

Question 17

Goal-Based Agent

Answer 17

The simple reflex agent and the model-based reflex agent do not have any goals. They just make decisions based on the current or past percept sequences. Yet the goal-based agent makes a decision with the percept-sequence and provides a distinction between a goal and non-goal state via an if-then statement. This provides us with a larger purpose, a target that we are trying to reach. This is an abstraction upon the goal-based agent.

Question 18

Task Environment

Answer 18

In designing an agent, the first step is always to specify the task environment as fully as possible. The task environment includes the following factors: PEAS.

P = performance measure 
E = environment (D-SOAKED)
A = actuators                       
S = sensors

Question 19

Utility Agent

Answer 19

The goal-based agent provides a crude distinction between reaching and not reaching a goal, that’s it. However, a utility-based agent lets us know precisely how close we are to our goal, we quantify how close the current state is to the goal.

The utility function is an internalized performance measure and can combine multiple conflicting goals. For example, the goals of a self-driving car are to not break the law and to get to a destination quickly. Because these goals conficlt, the utility function is a layer of abstraction and will specify the appropriate tradeoffs between the two.

Question 20

Types of Agents

Answer 20

These models are in order of increasing sophistication:
simple-reflex agent
model-based reflex agent
goal-based agent
utility-based agent

Question 21

Search Problem Definition

Answer 21

A search problem is a problem with a set of possible actions and resulting states and we must select the actions that will get us to the desired goal-state as efficiently as possible. A solution to this problem is defined as a list of all the actions one must take to get from the current state to the goal state.

This is the simplest kind of AI problem: episodic, single-agent, fully observable, deterministic, static, discrete, and known.

Question 22

Search Problem Components

Answer 22

State Space: the set of possible states this environment can be in
Initial State: the state the agent starts in
Goal State: the desired state
Actions: The set of actions available to the agent given a state s
Transition Model: a function that given a current state and an action returns the resulting state
Action Cost Function: the cost of going from one state to another via a certain action. This function maps (s, a, s’) to a real number cost.

Question 23

Best-First-Search()

Answer 23

The most basic algorithim to solve a search problem. This is the template for all other search-problem algorithms.