Demo Arcade Intelligence / JWIEDEMAN

Concept. Not built; this page describes a proposed design.

What it would do

Run a few simple arcade-style games (for example Pong or a maze game) on a cabinet or in a browser.
Let a visitor play against a reinforcement learning agent, or watch two agents play.
Retrain the agents between sessions on recorded games.
Show a leaderboard of human and agent scores.
Show, after each match, what the agent was paying attention to.

How it would work

The game runs at a fixed frame rate. Human and agent inputs go through the same input path, so neither gets a timing advantage.
During play the agent only runs inference. It does not learn mid-match.
Each match is logged: frames, inputs, scores.
Between sessions, the agent is trained further (for example with PPO) on new self-play games.
A new version replaces the old one only if it scores better on a fixed set of test games.

Parts

Part	Role	Candidate tools
Game	Rules, rendering, input	Browser canvas or a small game engine
Agent	Pick an action each frame	Small neural network policy
Training	Self-play and updates	Stable-Baselines3 or RLlib on one GPU
Match log	Store games for training and replays	Files or a small database
Leaderboard	Show scores	Static page plus a small API
Match summary	Show agent attention and rewards	Saliency maps, reward-over-time chart

Data it would need

Game environments with a clear reward (score, win or loss).
Self-play games for training. Human games are optional extra data.
A fixed set of test games to compare agent versions.

Risks and open questions

Agents often find bugs in the game and exploit them. Each game needs checks for this.
An agent that always wins is not fun. Difficulty may need to be capped or matched to the player.
Attention maps look convincing but can be misleading about why the agent acted.
A public leaderboard needs name filtering and a way to remove entries.
Accessibility: slower game speeds and remappable controls for human players.