Skip to content
View Nebraskinator's full-sized avatar
7 followers · 0 following

Block or report Nebraskinator

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
Nebraskinator/README.md

Zachary Ruhe

Molecular biologist (Ph.D., UCSB) who builds machine learning systems. I spent thirteen years at the bench on protein–protein interactions and bacterial toxin delivery, published ten first-author papers including one in Cell, then taught myself to program in 2020 and built the machine learning inside an FDA-cleared clinical diagnostic.

The repositories here are reinforcement learning work done on my own time, on a single consumer GPU. I design the architectures, write the training loops, and build the environments the agents learn in.

orbitwars — AlphaFold's Evoformer, applied outside biology

A Kaggle strategy environment turned out to have the same shape as a folded protein: entities, and the relationships between them. I adapted the Evoformer's node/pair message-passing — an explicit learned edge state, pair-biased attention, and tokens writing back into the edges each layer. Trained by imitation, then self-play PPO with a joint action distribution over the whole board.

ps-ppo — transformer reinforcement learning for Pokémon Showdown

Behavioral cloning, then distributed PPO across 250M+ states in two days on one RTX 3090. A custom attention mask isolates the state representation from the actor and critic heads, forcing the backbone to build an unbiased view of the game. 1900+ ELO from a pure neural policy — no tree search, expectimax, or damage calculator at inference. The published MLP baseline, reproduced under its own hyperparameters, plateaus near 1100.

SimpleTFTEnv — an environment built to isolate a problem

Simultaneous action selection, hidden information, and irreducible stochasticity, at a fraction of the compute. Built because the interesting failures in the Pokémon work needed a cheaper place to study them.

BFF — computational abiogenesis

Self-replicating instruction sequences emerging from random initialization in a 3D grid, under spatial constraint and simple interaction rules.

Interested in reinforcement learning, emergent behavior, and evaluation — specifically, how you tell whether a model is right for the right reasons.

Pinned Loading

  1. ps-ppo ps-ppo Public

    Transformer RL agent for Pokémon Showdown — 1900+ ELO from pure self-play, no tree search at inference

    Python 10 5

  2. BFF BFF Public

    Computational abiogenesis — self-replicating instruction sequences emerging under spatial constraint

    Python 4

  3. orbitwars orbitwars Public

    Evoformer-inspired transformer RL agent — node/pair message passing adapted from AlphaFold, trained by imitation then self-play PPO

    Python 3

  4. SimpleTFTEnv SimpleTFTEnv Public

    Multi-agent RL environment isolating simultaneous action, hidden information, and stochasticity

    Python 1 1

  5. StreetFighter2AI StreetFighter2AI Public

    Early RL experiment (2021) — Street Fighter 2 agent

    Python 1