agent.md

what is an agent?

• perceiving through sensor
• acting through actators

rationality

• the actions to be most successful -> measure success -> performance measure
• prior knowledge of env
• actions that agent can perform
• percept sequence

rational agent

• learning
• autonomy

Properties of task environments

• Deterministic: next state of the environment Completely determined by the current state and the actions
• Stochastic : otherwise
• Strategic environment: deterministic except for actions of other agents

structure of agents

• Agent = architecture + program

Share1: 惊喜最小化agent

• motivation: static and predictable status
• agent目的：长远来看环境熵最小，但不清楚完整的真实状态分布（需估计）

Share2: diffusion model

• motivation: 减少计算量，在低维空间训练

Types of agent programs

• Four Types
  • simple reflex condition-action-rules -> if...then... stated & status
  • model-based reflex far more complex condition-action-rules & dynamic
  • goal-based less efficient but more flexible & different goal -> different action
  • utility-based high-quality
  • *learning agent