Game Agent

What this event is about

Washington trial event for 2026–27. A team of up to 2 builds an original Godot 4.7.2 / GDScript game plus a simple Q-learning agent, documents the work in a 4-page Technical Design Report (TDR), then shows live training and defends the code in a 10-minute interview.

Game type options: side-scroller (obstacle avoidance), lunar lander, educational, or another type pre-approved by the Event Supervisor.

Official rules (hosted): Game Agent rules PDF. Season updates and extras live on the SciOly Game Agent support site.

Key ideas (with where to read them)

• 5-unit modular structure — Required architecture for your Godot scripts. Full definitions are in the rules under The 5-Unit Modular Coding Requirement. Short version:
  1. Core Game Logic — physics, obstacles, score, win/loss, rewards, and turning continuous values into integer states (Vector2i / Vector3i).
  2. Human Input Controller — keys only; no game rules or RL.
  3. Agent Brain — Q-table + ε-greedy math only (no graphics).
  4. Integration Loop — resets, episode count, wires Unit 1 ↔ Unit 3, training speed.
  5. Testing Harness — unit + integration tests.
• Technical Design Report (TDR) — A maximum 4-page engineering write-up (min 10-point font, single-spaced) submitted with the project. It replaces in-game docs. Required sections are listed under 4. The 4-Page Technical Design Report: unit-separation map, AI prompt/debug log, hyperparameter history, and human benchmark.
• Hyperparameter log — A dated text history of how you changed Alpha (α) learning rate, Gamma (γ) discount factor, and Epsilon (ε) exploration, and what that did to episodes-to-stable play. Background: Q-learning summary diagram and Applied AI Course Q-learning tutorial.
• Human performance benchmark — Your team’s documented human-player scores (method + results) used as the bar the trained agent should match or beat in the live interview. Coaches’ checklist language is in the rules PDF near the end; keep raw scores in the TDR from early practices.

Study plan

Setup (team of up to 2; Godot 4.7.2; digital submit Wed before competition; ~10 min live demo Saturday)

1. Install Godot 4.7.2, create a clean project, and bookmark the support site plus the WA Trial Event hub — scoring matrix, Q&A, and examples update there during the season.
2. Complete the recommended Muddy Wolf platformer tutorial (expect to rewind — the second half moves fast). Compare against the supervisor’s Example B game built from that tutorial.
3. Follow the official Video series outline (SciOly B) in order: splash screens → Python Q-learning intro → add the agent → run the optimal policy → troubleshooting. Use the Game Agent as built in the video series project as a reference (cleaner save/load of the trained policy).
4. Only after that path works, invent your game — still enforcing the 5-unit modular structure from day one. Unit 3 must not touch graphics; Unit 1 must not handle input. Rule of thumb from the video series: actions × states < 800.
5. Ship the required UI: splash/rules screen, Human Play / Fast Training / Agent Evaluation controls, live score, and WIN / LOSS / GAME OVER with a restart path (see competition interface rules in the PDF).
6. Keep a running hyperparameter log (α, γ, ε) and a human performance benchmark while you train — both become TDR sections and feed Live Interview scoring (TDR requirements).
7. Draft the ≤4-page TDR: (1) how you separated/tested the five units, (2) exact AI prompts + at least one AI-caused bug you fixed, (3) hyperparameter history, (4) human benchmark method/results. Cite any external sprites/SFX/libraries.
8. Practice the live loop: clear Q-table → train 2,000 episodes → run the learned policy → explain random lines of your code. Host target: i5-class, 16 GB RAM, integrated graphics, 1080p.
9. Submission hygiene (checklist in the PDF): delete .godot cache, zip cleanly, include TDR, email by Wednesday night. No OS.execute, no out-of-project file access, no runtime network downloads (instant DQ). Score: VPF/15 + AICH/20 + TDR/20 + RLMAP/25 + LI/20; first tiebreaker is Live Interview (scoring matrix).

Tips

Fancy graphics score less than clean modular code, honest AI debugging, and an agent that actually learns. Prefer the video-series Game Agent project over older example agents when copying save/load patterns. For RL vocabulary, start with Intro: AI and RML and the IBM reinforcement learning overview.