Skip to content

Latest commit

 

History

History
41 lines (38 loc) · 1.72 KB

project-ideas.md

File metadata and controls

41 lines (38 loc) · 1.72 KB

Project Ideas

Thoughts/Questions

  • Can GPT-4 play a good game of rock paper scissors?
    • could set up this experiment up through the OpenAI API
  • Can neuro-evolution lead to something that surpasses Iocaine Powder, in the sense that it is a very good general strategy?
  • Can neuro-evolution attain the level of Iocaine Powder purely through self-play?
    • If so, how large/complex would it need to be?
  • Is there an upper bound of complexity for RPS strategies?
    • Assuming not, would the evaluation period need to grow exponentially?
    • Also, if not, could RPS (or other generic task) act to "train" or "grow" a foundation model for time series in general?
  • Other than asymmetric rewards, are there ways of avoiding simplistic local minima?
  • Is there any relationship between RPS and compression algorithms?
    • Could this be a useful medium/input?
  • If we let evolution go on for a looooong time, do we ever observe things like double descent or grokking take place with generalization?
  • How well would a purely supervised learner do at this?
  • Does CMA-ES outperform SNES on this task?

Experiments

  • "Hall of champions" for improved self-play and generality.
  • More model types
    • Transformers
    • LSM/Echo State Networks?
  • Larger models
    • More hidden layers
    • Larger hidden layers
  • RL algorithms / online learning

Possible Features

  • Parallelization
    • helper function to setup Ray cluster on, say, AWS
  • Hyperparameter search
    • EvoTorch, Optuna, ...?
  • Named experiments, better config handling
    • SQLite?
  • Integration and unit tests
  • Web app leaderboard
  • GPU support
    • (current RNN models are too small for GPUs to help much)
  • More visualizations
    • Moving average of various n-grams through time used by agents