BVER: Bidirectional Voronoi-biased Exploration Curriculum for Reinforcement Learning

Iteration — 5 phases

speed iteration 0 goals reached 0 / 1

history live

Maze stage

Mode

Walk seeding — connect / voronoi

forward connect greedyness towards solved intermediate starts (blue dots)

backward connect greedyness towards solved intermediate goals (orange dots)

Policy simulation

skill transfer simulates how policy training on one expansion set affects the other

Envs & horizons

rollout envs

random-walk envs

random walk length

episode horizon longer trips time out

Import / export config

Load file

Export writes the current maze + all settings as JSON. To make one the boot default, paste it into the SHIPPED_DEFAULT constant near the top of this file's script.

Legend

Race: forward-only vs backward-only vs bidirectional

speed fwd connect 0.50 bwd connect 0.50 seeds averaged 5

connect ratios apply to the bidirectional pane; the single-direction panes force them to 0 (no opposite tree). Maze & env settings mirror the Explain tab — edit there, then return.

Forward-only

iter 0goals 0 / 1

Backward-only

iter 0goals 0 / 1

Bidirectional

iter 0goals 0 / 1

goals reached vs. iteration (mean over seeds)

covered region vs. iteration: fraction of free cells covered (mean over seeds)

Algorithm 1 — BVER (equation numbers refer to the paper)

Algorithm 2 — Random walk frontier expansion

Algorithm 3 — Return-based thresholding of rollouts into frontier and solved buffers

Algorithm 4 — Rebuild of the mixed set from frontier and solved buffers