REINFORCEMENT LEARNING · PHYSICAL DESIGN

Small circuits.
New possibilities.

Let an agent explore placement and routing.
Compare experiments. Replay the improvements.

Your machine.
Your experiment.Runs and results stay local.

EXPERIMENT NOTEBOOK

Pick up where you left off

Compare experiments

Only matching experiments Match the selected run, or the settings in the form.

Latest 10 matching runs.
Run / circuit	Seed / training	Steps · actual / budget	Grid → result	Area	Wires / crossings	Verified	First solution	Elapsed

01 / SET UP

Your experiment

Benchmark collection

Circuit

Technology

Clocking

Layout dimensions— tiles

Use reported minimum dimensions

A starting point from the published benchmark data, not a guaranteed result.

Width tiles

Height tiles

Training budgetPPO agent

Automatic timestep budget

Reading the input network to suggest a timestep budget…

Timesteps

Random seed

Consecutive seeds 1–10 runs

Timesteps per seed. Multiple seeds run sequentially on the same grid.

Stop after a verified solution Otherwise continue looking for a smaller, cleaner layout. Post-layout optimization Available with 2DDWave. Resume saved model Available after a model has been saved. Settings must match.

One agent at a time. Stop also cancels any remaining seeds; available results are kept.

02 / EXPLORE

Ready when you are

Ready

Choose a circuit and start your first experiment.

Training progress—

Timesteps—

Episodes—

Elapsed—

Best placed—

No experiment started yet.

Grid → optimized: —
Area · tiles: —
Wires / crossings: —
First verified solution: —

Complete placement and verified equivalence are reported separately.

Training statistics

PPO training epochs · this run: —
Including previous training: —

Training episodes are online exploration, not an independent policy evaluation.

Best observed placement—

Placement progress is available after training starts.

Mean episode reward

Rolling mean · latest 100 completed episodes

—

Start an experiment to see how the agent learns.

Rewards are measured after each completed episode. History is sampled for display.

03 / DISCOVER

A place for every gate

A little learning. A better layout.

The best observed placement will appear here.
This animation is an illustration, not a training result.

Ground layer Upper layerDrag to pan · scroll to zoom · arrow keys to pan · 0 to fit · click a tile for details

KEEP EXPLORING

Results & diagnostics

0 files

Available layouts, saved models and other outputs will appear here.

Training log