TriTon - A ternary neural network for managing Wi-Fi on OpenWRT routers

Community Builds, Projects & Packages
1

Hello OpenWrt community,

I wanted to share a project I've been developing and stress-testing on my own hardware (MIPS architecture without FPU, like MT7628).

TriTon is an autonomous, self-learning system for Wi-Fi optimization written entirely in pure POSIX Shell (Busybox/ash) with zero external dependencies.

Key Features:

  • Ternary Logic: Uses lightweight state evaluations (-1 / 0 / +1) based on integer comparisons instead of heavy float-based gradient descent. Perfect for low-resource routers.

  • Cascade Architecture:

    • Cascade 1 (Channel Selection): Parses iw scan via awk, evaluates neighbor APs (including HT40 overlaps), and dynamically chooses the cleanest channel.

    • Cascade 2 (Power Management): Monitors connected stations (iw station dump), tracks RSSI, retry ratios, and throughput, then fine-tunes TX power (5-20 dBm).

  • Reinforcement Learning: Saves state weights into /tmp/weights.conf to adapt to the local RF environment over time based on successful actions.

  • Stigmery (Multi-router Coordination): No central controller needed. Routers coordinate indirectly through environmental changes, automatically separating into non-overlapping channels.

In my real-world tests, it successfully reduced the average retry ratio to ~8.5% and stabilized throughput at minimal sufficient TX power (10 dBm), reducing interference for neighboring networks.

The project is fully open-source. I would highly appreciate your feedback, code review, or suggestions for optimization!

GitHub Link: https://github.com/den4ik86/TriTon

1 Like