Voyager by OutLauz

End Game with Circle of Death visible
Github action feedback showing vital game statistics after a push

Inspiration

For many complex problems, often machine learning approaches are thrown at them without much thought or understanding of the underlying processes. However, well thought-out heuristics are often better and more efficient approaches. With this in mind, we demonstrate that our approach, based on some novel heuristics, is able to outperform our competitors.

What it does

Voyager is an agent, optimised for playing SQLillo Royale, a Fortnite-inspired battle royale game. We are able to achieve top level performance using a number of sophisticated mathematical approaches such as exponential smoothing, linear algebra, offline planning and more. With these tools, we are able to develop efficient new algorithmic process for dodging mechanics, target analysis and situational intelligence.

How we built it

The main algorithms are implemented in Lua with JIT, analysis has been undertaken with Python, and we also make extensive use of Github actions for automation.

Challenges we ran into

Developing good heuristics requires intuition and deep knowledge of the game mechanics, in addition to the ability to formulate such heuristics as an optimisation problem to transfer to code. Another challenge was remote debugging without any logs since the code was running on a restricted remote server.

Accomplishments that we're proud of

None of us had any experience coding in Lua, so developing a highly performant algorithm for this game was a great accomplishment. We are also very proud of our Github actions pipeline we created, which streamlined the process of uploading our agents to the server, as well as calculating analytics and feedback on the agent's performance directly.