AMaze: fully discrete training with three regimes (direct, scaffolding, interactive) and two algorithms (A2C, PPO)

Dataset containing all training artifacts (final models, training curves, intermediate visualizations, ...) as well as the raw data used to assert generalization capabilities. The associated archive final_behavior.tar.gz provides a visualization of every replicate's final behavior for easier navigation. Distribution files contain a sampling across 1000 seeds, 5 probabilities for traps and lures, 4 sizes and 5 set sizes resulting in 486356 mazes. Descriptive graphs provide an overview of the accessible "maze space". v2: Added script to aggregate run dynamics (mean reward, errors, maze lengths...) and resulting generated dataset (csv) and plots (pdf)