How Reinforcement Learning Powers Front‑End Testing for Alibaba’s 618 Interactive Game

1. Interactive Intelligent Testing

The 618 interactive game introduced numerous state combinations, making manual testing costly. To generate realistic test data, the team applied reinforcement learning to simulate user behavior and used Puppeteer to run pages in a real browser environment for fast regression verification.

Reinforcement learning treats the problem as a Markov Decision Process (MDP) where an Agent interacts with an Environment , observes a State , selects an Action , and receives a Reward . The agent repeatedly performs actions such as collecting coins, buying carriages, and synthesizing carriages until it reaches the goal of winning a prize.

Agent: the decision‑making entity

Environment: the game world the agent interacts with

State: current status of the agent

Action: operation chosen based on the state

Reward: feedback from the environment

Snapshots of agent states are stored as test data for front‑end debugging. Example states are shown in the table below:

ID

Name

Carriage Info

Remaining Coins

State_1

Initial data

2 × carriage level 1

20000

State_2

Purchase carriage data

3 × carriage level 1

15000

…

…

…

…

2. Automatic Regression Testing

Front‑end interactions are complex, and traditional manual regression is inefficient. The team generated test cases automatically by replaying recorded user actions in a Puppeteer environment, selecting optimal actions via reinforcement learning, and building a user interaction behavior tree.

After constructing the behavior tree, redundant paths are removed, duplicate paths merged, and the remaining paths are split into concise test cases. Each test case is executed in Puppeteer, with UI snapshots compared using image‑recognition to verify component rendering.

3. Scalable Popup Production

The 618 popup count increased 2.5× compared to the previous year, demanding a faster development pipeline. To address this, the team decoupled the popup’s presentation layer from its logic layer.

The logic layer handles data models, initialization requests, and event handling, while the presentation layer receives data changes and renders UI accordingly. This separation allows the same logic to drive different UI forms (e.g., slot machine, wheel, or simple button).

A domain‑specific language (DSL) describes popup behavior, and a runtime engine interprets the DSL to generate the UI dynamically. The UI layer is treated as a pure function that consumes configuration and dynamic parameters, enabling rapid assembly of popup interfaces.

With a visual UI designer, developers can compose popup layouts from basic layers (group, image, text) and bind static or dynamic properties, drastically reducing UI development effort.

Overall, the intelligent testing solution and decoupled popup architecture became key tools for improving front‑end development efficiency and ensuring stable online operation during the 618 promotion.