Demo Arcade Intelligence

Project ID: AI-004 · Stage: Exhibition-ready showcase with continuous training

Mission Profile

Demo Arcade Intelligence is an experiential lab installation that pits reinforcement learners against human challengers in retro-inspired arcade environments. Each game cabinet streams telemetry to the lab, enabling policy iteration, interpretability experiments, and a living leaderboard.

Core Capabilities

Continuous training loop: Agents retrain after each tournament cycle using Proximal Policy Optimization with curriculum schedules.
Human drop-in mode: Visitors can instantly jump into the live environment; the system switches to inference-only mode while preserving fair scoring.
Explainability layer: Post-match briefings outline key decision branches, reward contributions, and input sensitivities for each agent run.

Systems Overview

Module	Purpose	Notes
Cabinet hardware	FPGA-based controller boards, 120 Hz displays	Deterministic latency pipeline for both human and agent inputs.
Training cluster	Kubernetes-managed GPU workers, Ray RLlib stack	Handles policy rollouts, evaluation, and checkpoint rotation.
Leaderboard service	Astro-powered microsite, Supabase backend	Publishes rankings, highlights hero runs, and archives telemetry.

Observability

Run cards: Automatically generated dossiers summarize score differentials, policy entropy, and notable events for each match.
Spectator HUD: Overlays agent attention heatmaps and reward accumulation so audiences can follow strategy shifts in real time.
Operator console: Allows lab staff to pause training, pin stable checkpoints, or trigger curated exhibition modes.

Safety & Fair Play

Agents undergo fairness checks to prevent glitch exploitation or soft-lock strategies.
Human sessions include accessibility presets such as slowed pace and remappable controls.
Leaderboard moderation rules ensure public handles remain appropriate and free from sensitive data.

Next Milestones

Add cooperative co-play scenarios where humans and agents collaborate toward shared objectives.
Release a public telemetry API for researchers interested in strategy evolution data.
Explore portable cabinet kits for traveling exhibitions and partner campuses.

Collaboration Signals

Program sponsor: Experience Design Lab
Primary engineer: Reinforcement Learning Architect
Contact: lab@jessicawiedeman.com

Document updated: 2024-05-12