Chapter 2: Intelligent Agents

• AI as the study of constructing systems that act rationally.
• More practical than the study of "rational thought"
• More general and easier to evaluate than the study of "human-like" thought and action.

Focus of this lecture:

• The definition of an agent.
• The interaction between agent design and the environment that the agent must operate in.
• The question of empirical validation.

Key terms:

• Architecture
• Perceptions --- sensors
• Actions --- effectors



• Program



• Environments:
• accessibility
• deterministic
• episodic
• static

What are agents:

• Agent = Architecture + Program



• Here we focus on the Program



• Types of Agents:
• reflex
• goal-based
• utility-based

Goal:

• Design agents that "do the right things," but what does this mean?

The question of empirical validation:

• Given options, the agent should choose those that are best for success



• Need an objective performance measure ---can't just ask the agent or its designer.



• Example criteria (may be conflicting):

• success at job?
• resources consumed?
• time taken?

The measure of success (rationality) must depend on knowledge; can't require omniscience.

What is rational at any given time depends on 4 things:

1.The performance measure

2.The percept sequence --- the complete perceptual history of the agent so far

3.Whatever built-in knowledge the agent has

4.The actions that the agent can perform

The ideal rational agent: does the action that is expected to maximizes the performance measure on the basis of the built-in knowledge and the percept sequence

• This can include doing actions to obtain more information

Autonomy:

• A system is autonomous to the extent that its behavior is determined by its own experiences



• It is reasonable to provide an agent with some initial knowledge as well as the ability to learn

Agent Design --- the agent program:

• An agent program maps from percept sequences to actions; the mapping may be simple or more typically very complex



• The architecture determines the range of possible percepts or actions



• The agent gets one percept at a time; the internal state can "remember" the past states, but not a total record (Why?)



• The goal or performance measure is not part of the skeleton program

Mapping from percepts to actions by table lookup:

• Uses entire percept sequence as an index into a table which contains the appropriate action



• problems with this approach:
• the size of the table
• no autonomy

static: percepts, a sequence, initially empty

table: a table, indexed by percept sequences, initially fully specified

append percept to the end of percepts

action <- Lookup(percept, table)

return action

Example: navigating while shopping

• reflex agents: stores floor plan pre-compiled in memory



• goal-based agent: creates an internal map; can adapt to unexpected changes



• utility-based agent: creates an internal map; can adapt to unexpected changes in a manner that maximizes the expected benefit

Simple Reflex Agent:

• condenses input sequence by noting commonly occurring patterns of input/output



• these patterns are expressed in terms of "conditional-action" rules
• e.g., if <percept> than <action>

• keeps no internal record of the state of the environment

static: rules, a set of condition-action-rules

state <- Interpret-Input(percept)

rule <- Rule-Match(state, rules)

action <- Rule-Action[rule]

return action

Reflex Agent with Internal State

• Maintains an internal state to distinguish between world states that generate the same percept but are different (because of past actions) with respect to the action to be taken.



• Requires the ability to represent change in the world

static: state, a description of the current world state

rules, a set of condition-action-rules

state <- Update-State(state, percept)

rule <- Rule-Match(state, rules)

action <- Rule-Action[rule]

state <- Update-State(state, action)

return action

Goal-Based Agent

• Contains some sort of goal information and knowledge about the results of possible actions performs the action or action sequence that achieves the goals



• See Fig 2.11 in text

Utility-Based Agent

• Goals designate desired states



• Utilities indicate preferences among states



• Particularly useful when:
• i. there are conflicting goals
• ii. no goal state can be achieved with certainty

• See Fig 2.12 in text

Examples of different types of agents: See Fig 2.3 in text

Connection to search and knowledge representation:

• Goal and utility-based agents require knowledge about:
• world states
• actions and effects
• goals and utilities



• Systems must reason with this knowledge; this process can be viewed as search

Environment properties (depends on how conceptualized)

1.Accessibility

2.Deterministic

3.Episodic vs. non-episodic

4.static vs. dynamic: will the world change while thinking

5.discrete: a finite number of percepts and actions

Accessibility?

• Accessible if sensors give all (relevant) information (if not accessible, then need some from the internal state "model")



• Example
• (accessible): chess
• (not accessible): medical diagnosis

Deterministic? The next state is completely defined by the current state and the agent's actions.

• An environment may seem non-deterministic when you just can't "see" what's effecting change; depends on Accessibility



• Example
• (deterministic): chess
• (not deterministic): dice

Episodic vs. non-episodic:

• Choices are local to each time slice; don't need to think ahead



• Example
• (episodic): robot part checker

Static vs. dynamic:

• will the world change while thinking
• Examples
• (static): chess, poker
• (dynamic): robot part checker, most of real life



• Semi-dynamic: world doesn't change but time costs

Discrete:

• a finite number of precepts and actions



• Examples: games, most "models" of real life

Summary of Agents

• 3 major types of agents:
• reflex
• goal-based
• utility-based



• the choice of agent depends on:
• the architecture
• goals
• the environment
• the performance measure



• Knowledge Representation: the representation of the information required by the agents; i.e., world states, actions and effects, goals and utilities



• Search: apply reasoning procedures to knowledge in "search" of the goal

Web site to check:

• MIT Media Lab

Homework:

• Read section 6.2 of your text describing the Wumpus world environment.

• Describe the Wumpus World environment in terms of the 5 properties listed above:

1. Accessibility
2. Deterministic
3. Episodic vs. non-episodic
4. static vs. dynamic: will the world change while thinking
5. discrete: a finite number of percepts and actions

• Discuss the appropriateness of each of the 3 major agent types for the Wumpus World:
• reflex agent
• goal-based agent
• utility-based agent