Pseudo-Code Algorithm Generator

2. Select Output Format

• LaTeX (default for papers/reports): Use algpseudocode + algorithm package unless user specifies algorithm2e
• Markdown: Plain-text pseudo-code blocks with consistent formatting

3. Apply Formatting Rules

These 10 rules apply to ALL pseudo-code regardless of domain:

1. One statement per line
2. Indent to show hierarchy (not braces)
3. Mathematical notation over code syntax (nabla, sum, argmax -- not for x in range(n))
4. Left-arrow for assignment (theta <- ..., not =). Equals is for equality testing only
5. Capitalize keywords consistently (Title-case: While, If, Return)
6. Number and caption algorithms ("Algorithm 1: Name")
7. List hyperparameters in Input/Require section -- never buried in loop body
8. Comments explain "why", not "what"
9. Consistent notation: bold for vectors (theta), calligraphic for sets/losses (L, D), greek for scalars (alpha)
10. Self-contained: readable from the algorithm box alone

4. Follow Domain-Specific Structure

Neural Networks -- Four canonical steps in the inner loop:

forward pass -> loss computation -> backward pass (backprop) -> parameter update

Early stopping sits between epochs. Validation is NOT inside the batch loop.

Reinforcement Learning -- Three-phase skeleton:

1. Initialize (parameters, value estimates, environment)
2. Loop (episodes or iterations):
   a. Interact with environment (collect experience)
   b. Compute targets (returns, advantages, TD errors)
   c. Update parameters (policy, value function, or both)
3. Return (policy, value function)

Use uppercase single letters for random variables (S, A, R). Advantage uses hat notation (A_hat).

Evolutionary Algorithms -- Population-level loop:

1. Initialize population
2. while not converged:
   a. Evaluate fitness
   b. Select parents
   c. Apply variation operators (crossover, mutation)
   d. Select survivors
3. Return best

Use (mu, lambda) notation. Set notation for population operations (union, empty set).

5. Avoid Anti-Patterns

• No programming-language syntax (loss.backward(), for x in range(n))
• No implementation details (data loading, logging, checkpointing, device management)
• No ambiguous = for assignment
• No missing initialization -- always show initial values before the loop
• No unlabeled algorithms -- always "Algorithm N: Name"
• No monolithic blocks >20 lines -- decompose into sub-procedures with \Call{}

LaTeX Quick Reference

algpseudocode (recommended)

\usepackage{algorithm}
\usepackage{algpseudocode}

\begin{algorithm}[t]
\caption{Algorithm Name}\label{alg:name}
\begin{algorithmic}[1]
\Require Input description
\Ensure Output description
\State $\boldsymbol{\theta} \gets \boldsymbol{\theta}_0$ \Comment{Initialize}
\While{not converged}
    \State $\mathbf{g} \gets \nabla_{\boldsymbol{\theta}} \mathcal{L}$
    \State $\boldsymbol{\theta} \gets \boldsymbol{\theta} - \alpha \mathbf{g}$
\EndWhile
\State \textbf{return} $\boldsymbol{\theta}$
\end{algorithmic}
\end{algorithm}

Key commands: \Require, \Ensure, \State, \If{} ... \EndIf, \While{} ... \EndWhile, \For{} ... \EndFor, \ForAll{} ... \EndFor, \Procedure{Name}{params} ... \EndProcedure, \Function{Name}{params} ... \EndFunction, \Call{Name}{args}, \Comment{text}.

algorithm2e (alternative)

\usepackage[ruled,lined,linesnumbered]{algorithm2e}

\begin{algorithm}[H]
\caption{Algorithm Name}\label{alg:name}
\KwIn{Input description}
\KwOut{Output description}
\While{not converged}{
    $\mathbf{g} \gets \nabla_{\boldsymbol{\theta}} \mathcal{L}$\;
    $\boldsymbol{\theta} \gets \boldsymbol{\theta} - \alpha \mathbf{g}$\;
}
\Return{$\boldsymbol{\theta}$}
\end{algorithm}

Key commands: \KwIn{}, \KwOut{}, \KwData{}, \KwResult{}, lines end with \;, \eIf{cond}{true}{false}, \If{cond}{body}, \While{cond}{body}, \For{cond}{body}, \tcc{multi-line comment}, \tcp{single-line comment}.

Critical: These two package families are mutually exclusive. Never load both.

Detailed Reference

For complete examples of every algorithm type (Adam, PPO, GA, CMA-ES, NEAT, Q-learning, etc.) with full LaTeX code, read references/algorithms-research.md.

Key sections to grep for:

• ## Neural Network Algorithms -- training loops, Adam, backprop, architectures
• ## Reinforcement Learning Algorithms -- Q-learning, SARSA, REINFORCE, PPO, A2C
• ## Evolutionary Algorithms -- GA, ES, CMA-ES, OpenAI ES, NEAT, GP
• ## Code Examples -- complete copy-paste LaTeX for Adam, PPO, GA, Q-learning
• ## Recommended References -- seminal papers and textbooks