← Workbench

Faber Neuroevolution · live workbench · EXP-051

Abandoning the objective: escaping a deceptive maze

Same start, same goal, same network. One search is rewarded for getting closer to the goal; the other for reaching new places, ignoring the goal entirely. On a maze built as a trap — the way through first leads away — only the second gets out. Every move here is the real faber-tweann engine on the server.

step 0 / 27

Chasing the goal

reward = get closer to G

G S

Seeking novelty

reward = reach new places

G S

Left, the goal-chaser climbs straight at the goal and jams against the wall one cell short — the cul-de-sac its own reward walked it into. Right, the novelty-seeker heads away, around the far side, through the distant gap, and reaches the goal.

The control: the same goal-chaser, on the twin maze

Identical maze, but the wall's gap is moved onto the greedy route. Now chasing the goal works — it solves in 8 steps. So it is the trap that stops it on the left, not the maze being hard.

Goal-chaser · twin maze

gap on the greedy route

reached ✓
G S

Now evolve it live

Run the real (μ+λ) evolution on the deceptive maze and watch the best-so-far path change each generation. Chase the goal and it converges into the cul-de-sac; seek novelty and it works its way around to the goal.

pick a reward, then evolve

more budget gives novelty more room to work around the trap

Seeking novelty

deceptive maze · live

G S

Deception is not a horsepower problem — in the signed experiment a strong sep-CMA-ES optimiser is trapped just as badly (0 of 40 runs), because a better goal-chaser only reaches the dead end faster. Changing what you reward, not the strength of the search, is what escapes.