Markov Decision Process: Continuous states

I wrote this post for lecture 17 in Andrew Ng's lecture collections on Machine Learning. In my previous post, we discussed Markov Decision Process (MDP) in its simplest form, where the set of states and the set of actions are both finite. But in real world application, states and actions can be infinite and even continuous. For example, if we want to model states of a self-driving car in a 2D plane, we must at least have the position \((x, y)\), the direction \(\theta\) of the car pointing to, its velocity \((v_x, v_y)\) and the rate \(r\) of change in \(\theta\). So the states of a car is at least a 6 dimensional space. For actions of a car, we can control how fast it goes in direction \(\theta\) and we can also control \(r\). Thus the actions have dimension 2.

In this post, we consider only continuous states with finite actions. Indeed, actions space usually has much lower dimension than states space, so in case of continuous actions, we might just discretize the actions spaces to get a finite set of representatives of actions. One may argue that we can also discretize the states space. Yes, we can do it, but only when the dimension \(n\) of state space is small enough: if we discretize each dimension into \(k\) parts, then there would be \(k^n\) many states. If \(n\) is large, then \(k^n\) is not feasible. This is so called the curse of dimensionality. Moreover, discretizing states space usually results in lack of smoothness.

Read More

Markov Decision Process

This article is my notes for 16th lecture in Machine Learning by Andrew Ng on Markov Decision Process (MDP). MDP is a typical way in machine learning to formulate reinforcement learning, whose tasks roughly speaking are to train agents to take actions in order to get maximal rewards in some settings. One example of reinforcement learning would be developing a game bot to play Super Mario on its own.

Another simple example is used in the lecture, and I will use it throughout the post as well. Since the example is really simple, so MDP shown below is not of the most general form, but only good enough to solve the example and give the idea of what MDP and reinforcement learning are. The example begins with a 3 by 4 grid as below.

the grid

Read More

The Phantom of the Opera


Read More

Principal Component Analysis

This article is my notes on Principal Component Analysis (PCA) for Lecture 14 and 15 of Machine Learning by Andrew Ng. Given a set of high dimensional data \(\{x^{(1)}, \dots, x^{(m)}\}\), where each \(x^{(i)} \in \R^{n}\), with the assumption that these data actually roughly lie in a much smaller \(k\) dimensional subspace, PCA tries to find a basis for this \(k\) dimensional subspace. Let's look at a simple example:

Read More




Read More

LeetCode Contest 60


第一题Flood Fill

给定一个二维数组表示一张图片,以及一个坐标(r, c)。我们需要包含这个坐标且数字一样的连通分支整体变成另一个数。

我记得高中的时候学到Flood Fill这个词的时候有种莫名开心,可能是因为这个名字很形象地描述了DFS的过程吧。

class Solution(object):
    def floodFill(self, image, sr, sc, newColor):
        :type image: List[List[int]]
        :type sr: int
        :type sc: int
        :type newColor: int
        :rtype: List[List[int]]
        visited = set()
        n = len(image)
        m = len(image[0])
        color = image[sr][sc]

        def dfs(r, c):
            image[r][c] = newColor
            visited.add((r, c))
            for dr, dc in [(0, 1), (1, 0), (0, -1), (-1, 0)]:
                x, y = r+dr, c+dc
                if x < 0 or y < 0 or x >= n or y >= m or image[x][y] != color:
                if (x, y) in visited:
                dfs(x, y)

        dfs(sr, sc)
        return image

Read More




Read More

Factor Analysis

This article is my notes on the topic of factor analysis. These notes come out of lecture 13 and 14 of Andrew Ng's online course. Roughly speaking, factor analysis models some \(n\) dimensional observed data with the assumption that these data are actually from some \(d\) dimensional plane in \(\R\), up to some Gaussian distributed errors. Let's make it more precise.

Suppose we have a set of observed data \(\{x^{(1)}, \dots, x^{(m)}\}\) implicitly labeled by some latent random variable \(z \in \R^d\) where

$$z \sim \mathcal{N}(0, I).$$

Factor analysis model tries to model \(P(x)\) using the assumption that

$$\begin{equation} x|z \sim \mathcal{N}(\mu+\Lambda z, \Psi), \label{cond-xz} \end{equation}$$

for some \(\mu \in \R^n, \Lambda \in \R^{n \times d}\) and diagonal matrix \(\Psi \in \R^{n \times n}\). These \(\mu, \Lambda\) and \(\Psi\) are parameters of the model.

Read More

Expectation-Maximization algorithm

In this article, I will collect my notes on Expectation-Maximization algorithm (EM) based on lecture 12 and 13 of Andrew Ng's online course. Given a set of unlabeled data points EM tries iteratively to determine the distribution of data, assuming that all data points are implicitly labeled (unobserved latent variables). For simplicity, we shall limit ourselves to the case where there are only finitely many implicit labels.

Description of the problem

Given a set of unlabeled data \(\{x^{(1)}, \dots, x^{(m)}\}\), our goal is to determine \(P(x)\), the distribution of \(x\), with the following assumptions.


  1. There are finitely many unobserved latent variables \(z \in \{1, \dots, k\}\) and they obey some multinomial distribution, i.e., \(P(z=j) = \phi_j\) with \(\sum \phi_j = 1\).

  2. \(\{P(x|z=j; a_j): j=1, \dots, k\}\) are a family of uniformly parametrized distribution.

Assumptions 1 and 2 will gives us a set of parameters \(\theta = (\phi_1, \dots, \phi_j, a_1,\dots, a_j)\) and

$$\begin{equation} P(x; \theta) = \sum_{j=1}^k P(x|z=j; \theta)P(z=j; \theta). \label{px} \end{equation}$$

We want to find this set of parameters so that the likelihood function

$$L(\theta) = \prod_{i=1}^m P(x^{(i)}) = \prod_{i=1}^m \sum_{j=1}^k P(x^{(i)}|z=j; \theta)P(z=j; \theta).$$

is maximized. Or equivalently, the log likelihood function below is maximized:

$$\begin{equation} l(\theta) = \sum_{i=1}^m \log\left(\sum_{j=1}^k P(x^{(i)}, z=j; \theta)\right), \label{log-likelihood} \end{equation}$$


$$P(x^{(i)}, z=j; \theta) = P(x^{(i)}|z=j; \theta)P(z=j; \theta).$$

Read More

Digit recognition, Softmax

在这篇博客中,我将使用softmax模型来识别手写数字。文章的第一部分是关于softmax模型的理论推导,而第二部分则是模型的实现。softmax的本质是一个线性模型,所以推导所需要的理论在我之前的一篇博客Generalized Linear Model已经详细介绍过了。softmax是逻辑回归(logistic regression)的推广:逻辑回归使用Bernoulli分布(二项分布),而softmax使用多项分布。

Read More