由于中间有很长一段时间在课上做ICS-PA去了，笔记中间断线了许多 虽然这课确实比较水没讲啥就是

Agent

Introduction

Agent: anything that can be viewed as perceiving its envivronment through sensors and acting upon that environment through actuators.

Rational agent: select an action to maximize given performance measure based on the percept sequence and built-in knowledge * Different from omniscience * Different from perfect

Designing agents: PEAS * Performance measure * Environment * Actuators * Sensors

Type of Environment

Fully observable: Everything the agent needs to make decision is accesible via its sensors. vs. partially observable, when not fully accesible, the agent needs to make informed guesses. In decision theory: perfect information vs. imperfect information.

Deterministic: agent does the change of world state. * The decision only depends on the agent decision and current state. vs. stochastic: the change of the world state beyond the control of agent (in some aspects). * Need to make guesses about the world change.

Episodic: current decision doesn’t depends on previous action. vs. sequential: current choice will affect future decision.

Static: the environment won’t change chronologically. vs. dynamic: the environment changes upon time. vs. semidynamic: the environment doesn’t change, but the agents performance score(e.g. the performance measure) will change with the passage of time.

Discrete: the percepts and actions is limited, dinstinct and clearly defined. vs. continuous: the percepts and actions is limited in a range.

Single agent: only one agent in the environment. vs. multiagent: multiple agent working together.

Type of Agent

1. Simple reflex agents percepts -> condtion-based action rules -> actions

   • Actions only depend on current percept.
   • Drawback: may develop deadloop.

2. Reflex agents with state/model

   • Know how world evolves.

3. Goal-based agents

   • Goal as the performance measure.
   • Know what it will be like if it does some action.

4. Utility-based agents

   • Consider the problem as Utilities beyond a simple goal.

• Learning agents
  • All the four agents above can be seen as a learning agents(if they get the ability to learn).
  • Has a problem generator, suggests exploring new actions when trying to solve new upcoming problems.

• State space graph: each state exists only once
• Seaarch Tree: same state can be in multiple nodes

Evaluating search algo:

1. completeness
2. time complexity -> can be evaluated by counting total nodes generated
3. space complexity
4. optimality

Aritificial Intelligence Beyond Classical Search

Local Search

Hill-Climbing Algorithm

Simulated Annealing

Local Beam Search

Genetic Algorithms

Constraint Satisfaction Problem

Logical Agent

Probability

茅庐川菜 湖区旁边市井川菜 陈麻婆豆腐 四牌楼外有鸡蛋灌饼 蓁巷里

Desicion Tree

信息熵

Regression and Classification

Linear Regression (线性回归)

parameter: 函数中要学习测定的变量

线性回归: 针对parameter而言

部分情况可以通过换元将问题转换到线性回归

loss function: 表示损失(精度)

e.g. square error function $\frac{1}{2m}\sum{m}{i=1}(h(x^{(i)}) - y^{(i)})^2$

通常的目标是最小化损失

Gradient Descent Algorithm (梯度下降法)

一个学习算法至少有两部分组成:

1. 损失函数
2. 优化算法

Regularization (正则化)

Logistic Regression (逻辑回归)

Aritificial Nerual Networks

Input: $x_n$

Parameters: $\omega_n$

Output: $a$

Activation function: $a = \sigma (\omega^T x + b)$

Procedure

1. Feed forward
2. Back propagation

Convolution Nerual Network (CNN)

kernel: 卷积核，是一个滤波器，在原有的图片上滑动并做卷积得到新的特征点阵

Transformer