Agent
preceiving its environment trough sensors
acting upon environment through actuators
Agent function
input: (sequence of) perception
output: action to take
Agent program
implementation of agent function
could be based on simple function table
can become impractical for complex problems, e.g. chess
Rational agent
agent funciton should maximize performance measure
rational agent != perfect agent
Challenge of AI
Function of rational agent should be encoded as “small” programm, not as a large function table.
Environments
PEAS:
Performance measure (e.g. safe/fast taxi ride)
Environment (e.g. streets, traffic lights, other cars, ..)
Actuators (e.g. steering, engine, …)
Sensors (camera, radar,…)
Properties of environments
Deterministic vs. stochastic
determinstic: next state of the agent is fully determined by the current state and the action executed by the agent
stochastic: otherwise
Fully vs. partly observable
Fully: Agent has access to the complete state of the env at each point in time (e.g. sensor can detect whole environment)
Partly: Agent has sensors to detect some part of the environment
Unobservable: No sensors
Episodic vs. sequential
episodic: future actions are not influcenced by current action (e.g. classification tasks)
sequentail: short-term actions have long-term consequences, usually agent stores internal state between decisions
Discrete vs. continuous
Depends on how time is handled:
discrete: chess, because there are a finite number of states
continiuous: taxi driving, because the process of driving involves that time passes continuously
Static vs. dynamic
dynamic: environment changes while agent runs decision function
static: otherwise
Single vs. multi-agent:
single: env only contains one agent
Types of agents
simple reflex
model-based reflex
goal-based
utitlity-based
Reflex-based
simple condition-action rules
e.g. initial vacuum robot
Model-based
similar to reflex-based
maintains internal model of the environment
action based on internal model and perception
Goal-based
agent has a goal
search for future states based on perception and goal
think ahead
can not be encoded as simple condition-action rules
Utility-based
maps state to utility measure (happiness)
Enviroinment, in which no simple reflex-based agent can behave rationally?
vacuum cleaner world
simple reflex-based agent does not maintain model of world
simple reflex-based cleaner can not memorize where to go in order to reduce number of steps
Last changed2 years ago
