论文写作 12: 算法伪代码 (含实例)

算法伪代码是论文的核心之一.

1. 需要说明输入、输出;
2. 方法 (函数) 名可写可不写, 如果被别的方法调用就必须写;
3. 需要写出主要步骤的注释;
4. 长度控制在 15-30 行;
5. 可使用数学式子或对已有数学式子的引用;
6. 不重要的步骤可以省略;
7. 一般需要进行时间、空间复杂度分析, 并写出配套的 property 以及相应的表格, 以使其更标准.
8. 在正文中分析伪代码的时候, 对多行引用应该用双连词符, 它会转成一个较长的连词符. 如: Lines 3–4 show. Tables 4–6 同理. 两个并行的词连接起来 (等价词的复合), 也应该用 Serial–parallel. 参考文献的页码也应该用双连词符 pp. 87–99. multi-label 这里有个从属关系, 就应该用单连词符.

例子:

以下是该算法的 tex 源码:

\begin{algorithm}[!htb]
	\renewcommand{\algorithmicrequire}{\textbf{Input:}}
	\renewcommand{\algorithmicensure}{\textbf{Output:}}
	\caption{Multi-label active learning through serial-parallel neural networks}
	\label{algorithm: masp}
	\begin{algorithmic}[1]
		\REQUIRE
			data matrix $\mathbf{X}$,
			label matrix $\mathbf{Y}$ for query,
            query budget $Q$,
            cold-start query budget $P$,
            number of representative instances $R$,
            instance batch size $B_i$,
            label batch size $B_l$
		\ENSURE
			queried instance-label pairs $\mathbf{Q}$, prediction network $\Theta$.
        \STATE Initialize the serial-parallel prediction network;
        \STATE $\mathbf{Q} = \emptyset$;\\
        // Stage 1. Cold start.
        \STATE Compute instance representativeness according to Eq. \eqref{equation: dp-representativeness};
        \STATE Select the top-$R$ representative instances to reorganize the training set $\mathbf{X}$;
        \STATE Update $\mathbf{Q}$ and $\mathbf{Y}'$ by querying $B_l$ labels for each of the top $\lfloor Q / B_l \rfloor$ representative instances;
        \STATE Train the prediction network using $\mathbf{X}$ and $\mathbf{Y}'$;\\
        // Stage 2. Main learning process.
		\REPEAT
            \STATE Compute $\hat{\mathbf{Y}}$ using the prediction network and Eq. \eqref{equation: label-prediction};
            \STATE Compute label uncertainty according to Eq. \eqref{equation: label-uncertainty};
            \STATE Query top-$B_i$ uncertain instance-label pairs to update $\mathbf{Q}$ and $\mathbf{Y}'$;
            \STATE Update the prediction network using $\mathbf{X}$ and $\mathbf{Y}'$;\\
		\UNTIL{($|\mathbf{Q}| \geq Q$)}
	\end{algorithmic}
\end{algorithm}

再来一个：

\begin{algorithm}[tb!]\caption{Active learning through clustering selection (TACS)}\label{algorithm: TACS}
  \textbf{Input}: A data matrix $\mathbf{X}$, cutoff distance ratio $rt$, the proportion of queried instances $qi$, the proportion of representative instances $qr$, small block threshold $ST$.\\
  \textbf{Output}: Queried/predicted labels $\mathbf{L} = (l_1, \dots, l_n)$.\\

  \begin{algorithmic}[1]
    \STATE $\mathbf{B} = [1..n]$; // Initialized as the block containing all instances
    \STATE $\mathbf{L} = (-1, \dots, -1)_n$; //Initialized as unknown
    \STATE $N_q = \lfloor n \times qi \rfloor$; //Number of queries
    \STATE $N_r = \lfloor N_q \times qr \rfloor$; //Number of representatives

	// Step 1. Select representative instances to label
    %\STATE Compute distance matrix $D = (d_{ij})_{n \times n}$;
    \STATE Compute $\rho_i$ for $1 \leq i \leq n$ according to Eqs. \eqref{equation: density-gaussian} and \ref{equation: d-c};
    \STATE Compute $\delta_i$ for $1 \leq i \leq n$ according to Eq. \eqref{equation: distance-to-master};
    \STATE $\gamma \leftarrow (\gamma _1, \cdots, \gamma_N) \leftarrow \rho \cdot \delta$; //Hadamard product for representativeness
    \STATE Query $N = N_r$ instances with the highest $\gamma_i$; // Current number of queries

    //Step 2. Query informative instances and classify
    \STATE $q$.enqueue($\mathbf{B}$); // The original queue with only one block
    \WHILE {($q \neq \emptyset$ \&\& $N \leq N_q$)}
        \STATE $\mathbf{B}$ = $q$.dequeue(); //Take a block from the head
        \IF {($|\mathbf{B}| \leq ST$)}
            \STATE continue; //Process small blocks later
        \ENDIF
        \STATE Query up to $\sqrt{|\mathbf{B}|}$ instances according to Eq. \eqref{equation: block-wise-representativeness} or \eqref{equation: distance-sum} and update $N$;
        \IF {($\mathbf{B}$ is pure)}
            \STATE For all $i \in \mathbf{B}$, $l_i$ = the label of queried instances in the block;
            \STATE continue; //No need to split further
        \ENDIF
        \STATE Cluster $\mathbf{B}$ to $(\mathbf{B}_1, \mathbf{B}_2)$ using the current best algorithm;
        \STATE $q$.enqueue($\mathbf{B}_1$);
        \STATE $q$.enqueue($\mathbf{B}_2$);
    \ENDWHILE
    \IF {($N < N_q$)}
        \STATE query $(N_q - N)$ unlabeled instances according to $\gamma_i$;
    \ENDIF
    \STATE Classify unlabeled instances using $k$NN;

    \RETURN $\mathbf{L} \leftarrow [l_i]_{n \times 1}$;
  \end{algorithmic}
\end{algorithm}