Classroom Activity: Guiding AI Through an EvoX Task via Dialogue

This page is a hands-on classroom activity: use a short conversational prompt to drive AI to generate and run a basic EvoX optimization task, then inspect the results the way an engineer would during acceptance testing. The point is not just getting it to "run"—it's validating"humans set direction, AI delivers the implementation, humans do the acceptance testing"as an AI Native workflow loop.

Activity Goal

Complete the minimal "prompt → generate → run → validate" loop in a real EvoX environment, rather than just producing readable code. You'll experience firsthand how to guide AI through a basic optimization task via dialogue, then inspect its output like an engineer.

Prompt Building Blocks (All Required)

Optimization problem & objective function: e.g., minimizing the Rastrigin function.
Search space & dimensions: e.g., 10 dimensions, with value ranges for each dimension.
Algorithm/operator selection: e.g., DE algorithm with crossover and mutation operators.
Population size & generations: e.g., population 128, 200 generations.
Compute backend: CPU or GPU tensorization.
Expected output: optimal value, convergence curve, run log.

Choosing Prompt Granularity

Prompts that are too vague trigger AI defaults, making it hard to judge whether your intent is met; prompts that are too detailed reduce the task to manual coding. We recommend specifying"constraints and acceptance criteria"and leaving implementation details to the AI. For example, specify the problem, algorithm, population size, and iteration count, but no need to dictate how exactly to call the API.

Run Results Checklist

Was it actually executed rather than just returning code? — Check the terminal logs and runtime.
Does the convergence curve trend monotonically better without abnormal oscillations? — The curve should decline steadily.
Does the optimal value match the known magnitude for the problem? — For example, the Rastrigin function's global optimum is 0.
Are the random seed and configuration reproducible? — Rerunning with a fixed seed should yield identical or similar results.

Follow-up Questions from an Acceptance Perspective

When results fall short, add constraints through dialogue—for example, "fix seed=42, rerun, and report the mean and variance across three repetitions." Observe whether the AI can iterate and refine rather than rewriting from scratch. This sharpens your ability to describe problems and acceptance criteria precisely.

Division of labor

Humans handle problem definition, metric criteria, and result judgment; AI handles API selection, code organization, and debugging execution. Through this activity, you'll gain a clearer sense of your role in the AI Native workflow.

Scope Boundaries

This page focuses only on validating the workflow for a basic task; it does not cover follow-up topics such as MetaDE bi-level evolution, multi-objective trade-offs, or Genesis multi-agent systems.

Sample Prompt

"Use EvoX to minimize the Rastrigin function (10 dimensions) with the DE algorithm, population 128, 200 generations, GPU tensorized backend, fixed seed=42; please run it directly and return the best fitness, convergence curve plot, and run log."

During acceptance, check four things: Was it actually executed? Is the shape of the convergence curve reasonable? Is the optimal value close to 0? Are repeated run results consistent?

EvoX Classroom Activity Loop

This interactive screenshot shows the minimal classroom activity loop: use a prompt to have AI generate and run an EvoX optimization task. The prompt must specify the problem, objective function, search space, algorithm, population size, generations, compute backend, and expected output. After running, validate by checking whether it was actually executed, whether the convergence curve is reasonable, whether the optimal value is close to the known magnitude, and whether a fixed seed produces reproducible results—demonstrating that humans set direction, AI implements, and humans validate.

Question 1

In this classroom activity, when constructing a conversational prompt, which of the following is NOT a required element to specify?

Population size & generations

The exact number of code lines for the algorithm implementation

Optimization problem & objective function

Search space & dimensions

Question 2

When checking run results, which of the following is key evidence that verifies "actual execution" rather than just generated code?

The terminal output includes iteration logs and a convergence curve plot

The code has detailed comments

It uses a GPU backend

The code contains the correct function names

Question 3

When AI's optimization results fall short of expectations, what is the most effective next step?

Completely rewrite the prompt and start over

Manually edit the generated code

Add constraints through dialogue, such as fixing the random seed and rerunning

Abandon the task and switch to another algorithm

Question 4

In the "humans set direction, AI implements, humans validate" loop, what are humans mainly responsible for?

Tuning algorithm hyperparameters

Writing and debugging code

Selecting specific APIs and libraries

Define the problem, metric criteria, and result judgment

Question 5

Regarding prompt granularity, which approach is most appropriate?

The vaguer the prompt, the better—let AI improvise freely

Specify constraints and acceptance criteria clearly, leave implementation details to AI

Give only the problem name, leave everything else as defaults

The more detailed the prompt, the better—equivalent to manual coding

Case Study: Guiding AI Through an EvoX Optimization Task via Dialogue

Read the following scenario and answer the questions.

An engineering team uses dialogue to guide AI through an EvoX optimization task: minimizing the 20-dimensional Ackley function with the CMA-ES algorithm, population 64, 500 generations, CPU backend, fixed seed=123, requesting the optimal value and convergence curve. After running, they find the convergence curve oscillates violently in the early stage, and the final optimal value is far greater than 0 (the Ackley global optimum is 0). The team wants to diagnose and fix the problem.

Analyze possible causes of the convergence curve oscillations and abnormal optimal value, and propose at least two methods for diagnosing or correcting them by adding dialogue constraints.

Answer:
Possible causes: step-size parameters too large, insufficient population diversity, algorithm parameters mismatched with problem dimensionality, random seed causing early-stage exploration fluctuation, etc. Constraint-adding methods: ① Fix the seed and rerun three times, computing the mean and variance to observe result stability; ② Request output of the best value change per generation to check for convergence; ③ Adjust the step size or population size and rerun.

During the acceptance stage, besides checking the optimal value, what other aspects should be examined to ensure reliable results? List at least three checkpoints and explain their significance.

Answer:
① Actual execution: Check terminal logs and runtime to ensure the code was actually run, not just generated. ② Convergence curve shape: Whether it improves monotonically and whether there are abnormal oscillations, to judge algorithm stability. ③ Optimal value magnitude: Compare with the known global optimum of 0; if far greater than 0, it may not have converged. ④ Reproducibility: Rerun with a fixed seed; results should be consistent, otherwise randomness is interfering. ⑤ Reasonable runtime: Excessively long runtimes may indicate efficiency issues.