CSE 471598, CBS598 Intelligent Agents

1
CSE 471/598, CBS598Intelligent Agents
TIP Were intelligent agents, arent we?
2
Introduction

• An agent is anything that can be viewed as
  perceiving its environment through sensors and
  acting upon that environment through actuators.
• We discuss
• A wide variety of agents
• How to define an agent
• Lets look at Figure 2.1
• Is that me?
• An agent function maps percepts to actions

3
All about Agents

• We will learn
• How agents should act
• Environments of agents
• Types of agents
• human, robot, software agents
• A vacuum-cleaner world with 2 locations (Fig 2.2)
  
• Percepts location and contents, e.g., A,Dirty
• Actions Left, Right, Suck, NoOp
• A simple agent function(Fig 2.3)
• What makes an agent good or bad?
• We need to specify how agents should act in order
  to measure

4
How Agents should act

• A rational agent is one that does the right
  thing.
• What is right? The issue of performance
  measure, is not a simple one
• You often get what you ask for.
• Be as objective as possible
• Measure what one wants, not how the agent should
  behave
• E.g., how to be a rational instructor/student?
• Do the right thing and be objective. How?
• Switch your positions (student, instructor,
  another student)
• A related issue is when to measure it.

5
A rational agent is not omniscient

• Rationality is concerned with expected success
  given what has been perceived
• A percept sequence contains everything that the
  agent has perceived so far
• An ideal rational agent should do whatever action
  that maximize its expected performance
• Rationality does not mean perfection which
  maximizes actual performance
• Do we sometimes regret? Do regrets help? Why?

6
Four key components

• What is rational depends on PEAS
• Performance measure
• Environment
• Actuators generating actions
• Sensors receiving percepts
• Another example?
• Taxi driver
• Lets look at its performance measure
• Refer to Fig 2.4

7
Definition of a rational agent

• For each possible percept sequence, a rational
  agent should select an action that is expected to
  maximize its performance measure, given the
  evidence provided by the percept sequence and
  whatever built-in knowledge the agent has.

8

• From percept sequences to actions
• A mapping with possibly infinite entries
• An ideal mapping describes an ideal agent
• Its not always necessary to have an explicit
  mapping in order to be ideal (e.g., sqrt (x))
• An agent should have some autonomy.
• i.e., its behavior is determined by its own
  experience.
• Autonomy can evolve with an agents experience
  and percept sequence - learning.

9
External environments

• Without exception, actions are done by the agent
  on the environment, which in turn provides
  percepts to the agents.
• Environments affect the design of agents
• Types of environments

10
Types of Environments

• Fully vs. partially observable
• Deterministic vs. stochastic
• E is deterministic but actions of other agents
  are not strategic
• Episodic vs. sequential
• An example of episodic environment?
• Static vs. dynamic
• E does not change, performance score does
  semi-dynamic
• Discrete vs. continuous
• Single vs. multiple agents
• What is the most difficult environment?
• Lets look at some examples in Fig 2.6

11
Design and Implementation of Agents

• Design an agent function that maps the agents
  percepts to actions.
• Or to realize how actions are selected/determined
• Implement the agent function in an agent program
  which is realized in an agent architecture
• Agent Architecture Program
• Percepts and actuators function mappings
• From Robots to Softbots (Amazons A9)
• Architectures can be very different

12
Some examples of agents

• All agents have four elements (PEAS)
• 1. Performance 2. Environment
• 3. Actuators 4. Sensors
• Fig 2.5 demos some agent types
• We can see that there are many ways to define
  these components and its impractical to
  enumerate all possibilities

13
Starting from the simplest

• A look-up agent (Fig 2.7)
• Generates actions based on percept sequences
• Your decision today is determined by many things
  happened in the past
• Why not just look up?
• How far back should we look up
• Scaling up
• An equivalent question is about the table size
• What else should we try?

14
(No Transcript)
15
Types of agents

• Simple reflex agents - respond based on the
  current percept, ignore the percept history. It
  cuts down a lot of possibilities.
• An example (Fig 2.8)
• A simple reflex agent (Figs 2.9,2.10)
• Condition-action Rules
• Innate reflexes vs. learned responses
• What if the environment is not fully observable?

16
Model-based reflex agents

• They can handle partial observability
• Knowledge about how the world works is called a
  model of the world
• Maintain internal state to keep information of
  the changing environment and involve
  consideration of the future
• Respond to a percept accordingly (Figs 2.12)
• From local to global

17
(No Transcript)
18
Goal-based agents

• They aim to achieve goals
• Goal desirable states,
• Search for a sequence of actions,
• Plan for solving sub-problems with special
  purposes
• Goals alone are often not enough to generate
  high-quality behavior. Why?

19
(No Transcript)
20
Utility-based agents

• They aim to maximize their utilities
• Utility the quality of being useful, a single
  value function
• Happy or not (a goal or not) vs. How happy when
  the goal is achieved
• resolve conflicting goals (speed vs. safety)
• evaluate with multiple uncertain qualities
• search for trade-off facing multiple goals

21
(No Transcript)
22
Learning agents

• They can learn to improve
• Operate in initially unknown environments and
  become more competent
• Four components (1) problem generator (to create
  exploratory actions), (2) performance element
  (the earlier entire agent), (3) learner, (4)
  critic (to provide feedback)
• The above types of agents can be found in the
  later chapters we will discuss.

23
(No Transcript)
24
Summary

• There are various types of agents who cannot live
  without external environment.
• Efficiency and flexibility of different agents.
• Using ourselves as a model and our world as
  environment as a starting point (Are we too
  ambitious?), you may
• Describe options for future consideration
• Recommend a new type of agents