Devsisters: Reinforcement learning agent for stage testing in Cookie Run: Witch's Castle

Problem

Devsisters needed to test and balance stages for Cookie Run: Witch's Castle. Some bugs only appear after thousands or millions of rounds, which manual play cannot reach.

Approach

The in-house Data Science team built a reinforcement learning system, called Nimbus, in which an agent plays the game. How well the agent does guides changes to stage difficulty, and the agent also finds bugs. The team used Weights & Biases to track and debug training runs. The system supports a two-week stage update cycle.

Results

2-3 days to 3-4 hoursdevelopment cycle length for some developers, per Weights & Biases' customer storyVendor case study[1]

50 stagesstages the system finished training in a single day, per Weights & Biases' customer storyVendor case study[1]

Takeaway

An automated player gave designers fast feedback on difficulty and bugs. The speed figures come from the tool vendor's customer story and quote only some developers.

Sources

Devsisters Levels Up Game Development with W&B
Weights & Biases2023Vendor publication

Source summaries

[1]Weights & Biases2023

Weights & Biases' customer story describes how Devsisters' Data Science team built a reinforcement learning game tester for Cookie Run: Witch's Castle. It says some developers saw a development cycle fall from 2-3 days to 3-4 hours and that training for 50 stages finishes in a day.

Full details are in the original sources.