Lecture 2

Announcements

• Office hour changed to 10-11am tomorrow
• Seminars (every tuesday 3725BBB 4:10-5:15)

What are agents?

Goal: How to build these types of agents?

• what about the human agents? they run on behave of somebody.
  • sport agents, actor agents, lawyers, contractors, travel agents, secret agents
• possible examples of computer-based agents?
  • watson: for medical setting instead of jeopardy setting (?)
• why we call them "agents"? instead of programs?
  • keep up with the latest buzzwords.

digraph agents {
  "Environment" -> "Agent" [label="precepts"];
  "Agent" -> "Environment" [label="actions"];
}

An Agent program ($\pi$): A real-time computable function from observations to actions.

$\pi^t: O^{t-1}\to A$ $a_t = \pi^t(\langle o_1, \cdots, o_{t-1}\rangle)$ $\pi = \langle \pi^1, \pi^2, \cdots \rangle$

Evaluating agent behavior

examples: spam filter, stock money, crashing into things on driving?

• what do we expecting on the agents?

PEAS Description

• P: Performance measure: 要如何衡量成效
• E: Environment: 環境給我們什麼
• A: Actuators: Agent 可以如何影響環境, 可以做什麼?
• S: Sensors: 要如何測量/感知周圍的環境, 可以做到多好?

Example: Google Car

• Performance
  • Get to a place fast or get to a "right" place fast
  • Confidence or Safety
  • profit
• envirnoment limit: weather, highways
• Sensors: wheels, distance sensors, ...
• Actuators: making the engine go, breaking system, controllers...

Environment

• Fully vs Partially observable
  • ex: chess
• Deterministic vs stochastic (隨機地)
• Episodic vs Sequential (是否循序累積?)
• Static vs Dynamic
• Discrete vs Continuous
• Single-Agent vs Multiagent
• known vs unknown

Example: High-frequency trade agent

Rational Agents

Does not require omniscience.

Could involve information gathering

The role of coherence

• Abstract specification of a large class of behaviors.

Agent designs

• Model based agents
  • a set of paper, rock, scissers
  • ruled based agents
• Goal based agents
  • if goal changed, keep tracking on
• Utility-Based agents
  • pay a dollar

Evaluating agent designs

• Analytical proofs: prove that there is no better program
• Empirical evaluation

possible world representations

• Atomic
• factored
• structured