Arcade Lab

Arcade Lab#

Arcade Lab is a laboratory for watching a search process happen in public. The browser runs many complete copies of an arcade game. Each copy—called a walker—chooses discrete actions, advances for a small number of frames, records what it earned, and may later be copied by another walker. The pictures on the screen show the current evidence produced by the search; repeated trials accumulate the evidence that guides the swarm.

Open Arcade in the browser. The page puts Mario, Atari, Sonic, and Montezuma behind one interface. It runs the Fractal Gas machinery and emulator cores as WebAssembly; parallel workers let a crowd of game copies move together. The useful habit is to watch the swarm and the reward plots at the same time. A single exciting score can be luck. A population that repeatedly finds and preserves progress is evidence of a different kind.

Arcade Lab sits next to the continuous-control and optimization laboratories. Here the state is a discrete emulator snapshot, while smooth-surface models use points; a one-frame mistake can change everything.

Open the browser laboratory#

From the repository root, serve the browser build with the command below, then open the local address. The small server supplies the cross-origin-isolation headers required by SharedArrayBuffer and WebAssembly threads. Directly opening index.html and ordinary static serving provide the page itself, while this command adds the headers that enable the threaded runtime. This command serves the existing fractal-gas-web/web assets. Rebuilding the WebAssembly artifacts follows the Emscripten workflow in fractal-gas-web/README.md.

make web
# open http://localhost:8000/web/arcade.html

Note

On a hosted build, the ROM vault may ask for one password. The browser decrypts the bundled ROM locally and remembers the successful unlock in that browser; the ROM remains inside that browser. Sonic also accepts a one-time local upload of your own Genesis ROM. Local development can use the plaintext ROM files when they are present.

The mental model: a crowd of timelines#

Imagine putting 48 identical game machines on a table and giving each one a slightly different sequence of button presses. After a short while, some machines have moved farther, found a ring, entered a new room, or simply stayed alive. Arcade Lab does the bookkeeping for this crowd. A walker carries the emulator state, its reward history, and any game-specific information needed to continue from that exact moment. The map shows one projection of this richer object.

Now comes the important operation: cloning copies a complete timeline—the emulator state, reward history, and game-specific context together. A promising walker can become the starting point for another attempt, while the other attempts continue to explore. In Wave, the population stays fixed and every walker steps on each iteration. In Graph, visited states remain as nodes in a growing tree, and active leaves are the states from which new alternatives can be extended. The graph is therefore a record of alternatives that preserves its earlier branches.

FMC and Jump Wave use the same search population differently: they look ahead, then advance one committed game and search again from the resulting emulator snapshot. That distinction matters when reading the screen. Wave and Graph show a leading search walker; the planners show the game being played, while the plots still describe the search population.

Four environments, four kinds of difficulty#

The four buttons open four benchmark setups, each with its own state variables, reward signals, and map construction. The defaults are chosen to make the state useful to the swarm: Mario, Sonic, and Montezuma start in Coords mode, while generic Atari starts in RAM mode because its generic interface uses a shared memory representation across the game picker.

Setup	Browser runtime	Default observation and reward	Map in the app
Mario	NES, `nes-py`; World/Stage selectors	Coords; shaped progress with time, death, area, and flag terms	Full level map with the swarm overlaid
Generic Atari	Atari 2600, ALE; bundled game picker, default Ms. Pac-Man	RAM; the selected game’s own score	Gameplay view with the selected game’s score; generic Atari allocates this display to gameplay
Sonic	Sega Genesis, Genesis Plus GX; Zone/Act selectors	Coords; shaped progress through the act, rings, score, lives, and completion	Fog of war assembled from walker views
Montezuma	Atari/ALE with room-aware logic	Coords; score plus a new-room bonus	A 24-room pyramid assembled as rooms are found

The table also tells you why raw reward values deserve game-by-game comparisons. Mario’s reward uses shaped progress, Atari reports its own score, and Montezuma’s room bonus is an explicit exploration signal. Likewise, maps take different forms: some use a supplied level image, Sonic’s map is stitched from the swarm’s downsampled views, and Montezuma’s rooms appear as walkers discover and render them. In both cases, unrevealed territory marks portions the swarm has yet to see, while those portions remain part of the game level.

Controls that change the experiment#

Start with Wave. Each iteration gives every walker a random action and holds it for dt frames, with dt drawn uniformly between dt min and dt max (the defaults are 6 and 30 frames, roughly 0.1 to 0.5 seconds at 60 frames per second). The cloning step uses virtual reward: a combination of how well a walker has done and how far it is from its comparison walker, with an optional visit-count term. Thus a good score and a good virtual reward are related, while they remain distinct measurements.

Observation chooses what the distance calculation sees: console RAM, RGB pixels, grayscale pixels, or a compact Coords tuple. It preserves the game’s controls. Changing observation, algorithm, game, or start level restarts the run because these choices change the state that is being copied. The seed, walker count, elite buffer, fitness coefficients, visit settings, and frame range are the experimental knobs; keep them with a result you want to repeat.

The reward-term sliders for Mario, Sonic, and Montezuma apply to rewards earned from that point onward. Reward already banked by a walker keeps its old weighting, so a live change leaves its history intact. Pause preserves the current search or trajectory. Reset reconstructs the run with the same settings and seed. The Env frames readout counts frames actually emulated, including each action step that concludes with death or life loss.

Warning

Read the Virtual reward plot as the cloning signal alongside the game’s score. Virtual reward is the selection signal used for cloning; Cumulative reward is the reward accumulated by each walker. A swarm can improve the selection signal by spreading out, while a game score can remain flat. Read both plots together.

Follow the Arcade Lab guide#

Use the short guide pages when you want to turn the picture into an experiment:

Getting started with the Arcade Lab: open the app, run the first swarm, and read the screen, map, plots, and status values.
Arcade Lab environments: learn the game-specific observations, rewards, start-level controls, maps, and termination behavior.
Arcade Lab controls, settings, and diagnostics: work through algorithms, planning, observations, fitness, visits, frame skips, seeds, and live reward settings.

For a first session, choose Mario in Coords mode, leave Wave selected, press Start, and watch how the dots and plots change together. Then switch to Graph and ask what the retained branches reveal; after that, try a planner and notice the difference between searching and committing an action. That sequence makes the controls answer a concrete question instead of becoming a collection of knobs.