Intelligent Agents
Last updated
Last updated
Agent 可以是任何一種東西，用 censor 來感測環境，並且透過 actuators 來動作
An agent’s behavior is described by the agent function that maps any given percept sequence to an action.
e.g., vacuum-cleaner
if the current square is dirty, then suck
otherwise, move to the other square
Rational agent 指的是可以做出正確事情的 agent
正確事情 : Agent 所做的事情會讓 enviroment 歷經一系列的 desirable states
desirability : performance measure 來量測
Definition : Rational agent 會在 percept 的每個步驟都需要 maximize performance measure，從而 action
Example : vacuum-cleaner agent
Performance measurement : the amount of dirt being cleaned up
Performance measure: a clean floor
Rationality maximizes expected performance, while perfection maximizes actual performance.
針對 Agent 的 rationality 不需要達到 omniscience
但 Agent 需要能夠從 perceive 的 data 中學習 (learn)
另外 Agent 最好能夠 Autonomous
learn and compensate for incorrect
我們透過以下幾點，來定義一個 Task environment :
Performance measure
The environment
Agent's actuators
Agent's sensors
合稱為 PEAS (Performance, Environment, Actuators, Sensors)
例如 :
Agent 的每個 action 是否會跟 sensors 偵測所有的東西有關
如果有多個 Agents，那他們之間是否會互相影響對方
Chess 就是 multiagent
若 environment 的下一個 state 會受前一個 state 影響，那就是 deterministic
stochastic 則相反，通常代表 outcomes 是 uncertainty 的
Taxi driving 就是 stochastic
情節之間是否是連貫的
Chess & Taxi driving 都是 sequential
Environment 會因為 Agent 的思考而改變，則說 environment 是 dynamic 的
Taxi driving 是 dynamic 的
Chess 的 environment 有 finite 的 discrete states & actions
Taxi driving 則是有 continuous state
The agent's state of knowledge about the “laws of physics” of the environment.
Agent = Architecture + Program
Program : 設計來 implement agent function (用於 map percepts to action)
Architecture : 指的是 Devices (sensor, actuator)
A trivial agent program turns percept into action each time :
下一個 action 只建立在 current percept, 而會無視以前的 percept history
只建立在 environment 是 full-observable
Agent 會持續追蹤外部世界的變化
並更新內部的 internal state
Rule 跟 Simple Reflex agents 的是一樣的
因為只有 current state of environment 沒辦法每次都能產出好的 action
所以基於 model 再新增 goal information
Behavior 可以因 Goal 改變而被簡單改變
Decision making 多了考慮未來的部分 "What/will"
有 Goal 不見得能有 High quality behavior
Utility-based agent 會挑選能夠 maximize expected utility 的 action
使用 critic 的 feedback 來決定 performance element 要如何修改來加強 action
learning element => making improvements
performance element => selecting actions
learn and improve in short run => better actions in long run
problem generator suggest new actions
Agent
Performance
Environment
Actuators
Sensors
Taxi driver
Safety, fast
Traffics, Customers
Brake, Accelerator
GPS, Engines
