Autonomous Strategy Research Game (Working Title)

A Game Design Document (Draft 0.1)

1. Overview

This project is a hybrid between a video game and a research simulator.
The core goal: enable autonomous AI agents to compete with one another using different strategies, doctrines, and AI paradigms, under limited compute budgets.

Unlike traditional RTS/MOBA titles, this game emphasizes:

Long-form strategic interactions condensed into short simulations (seconds–minutes per match).
Multiple viable strategies (e.g., swarm, attrition, economy control).
Comparative analysis of AI decision-making approaches.

2. Core Pillars

Autonomy: AI players control everything — scouting, resource collection, combat, and mission execution.
Strategic Diversity: Different doctrines (swarm rush, attrition, economy, hybrid) must be viable.
Compute as a Resource: Players must budget AI complexity vs. raw unit count.
Research-Oriented: Outputs are structured to allow analysis of AI behavior and strategy.

3. Game Structure

Ticks: Game proceeds in discrete time units (“ticks”).
Duration: ~5000 ticks per match (1–2 minutes real-time simulation).
Win Conditions: Based on missions/objectives, not just annihilation.
Asymmetry: AI paradigms may differ per player (NNs, FSMs, symbolic AI, hybrids).

4. Core Gameplay Loop

Setup Phase
- Player (human designer) configures units, towers, and AI controllers.
- System enforces compute budget cap.
Simulation Phase
- Each tick, active AI controllers process game state and issue orders.
- Resource collection, unit production, and combat resolved automatically.
- Fodder units follow orders with minimal compute overhead.
Resolution
- Match concludes after 5000 ticks or objective completion.
- Results recorded for analysis (win/loss, efficiency, resource usage, strategy path).

5. Compute Budget System

Each player has a fixed compute budget (e.g., 100 points).
They must “buy” AI control power for towers, squads, and special units.

Example Costs

| AI Type | Cost (per instance) | |—————————|———————| | Neural Net (small) | 10 pts | | Neural Net (medium) | 20 pts | | Neural Net (large) | 30 pts | | Symbolic AI | 8 pts | | FSM / Rule-based AI | 3 pts | | Squad Controller (shared) | 12 pts | | Hero Unit (w/ NN) | 15–25 pts | | Dumb Fodder (no brain) | 0 pts |

Example Builds

Swarm (Zerg-like): Many fodder units controlled by few squad brains.
Attrition (Turtle-like): Heavy defenses, fewer but smarter units.
Hybrid: Balanced squads, economy control, selective heroes.

6. Units & Structures

Fodder Units: Cheap, controlled by squad AI, no individual brains.
Squads: Groups of units sharing one AI brain (saves compute).
Hero Units: Expensive, individually controlled by powerful AI.
Towers: Defenses or production facilities, can host symbolic/FSM/NN controllers.

7. AI Controllers

Abstracted interface allowing different model types:

Neural Networks (NNs) — adaptable but costly.
Finite State Machines (FSMs) — simple and efficient.
Symbolic AI / Rule-based — interpretable strategies.
Hybrid Controllers — mixing methods within a player’s build.

8. Match Flow (5000 Ticks Example)

0–500 ticks: Scouting & early economy.
500–1500 ticks: Initial skirmishes, resource claims.
1500–3000 ticks: Midgame — tech upgrades, squad maneuvers.
3000–4500 ticks: Attrition battles, positioning, harassment.
4500–5000 ticks: Endgame pushes, win conditions resolved.

9. Research Applications

Strategy Evaluation: Measure effectiveness of doctrines under constraints.
AI Comparisons: Test symbolic vs. neural vs. hybrid approaches.
Autonomous Warfare Modeling: Explore multi-objective missions and resource tradeoffs.
Data Output: Structured logs for replay, analysis, and training datasets.

10. Future Development Goals

Full design of unit cell system (custom modular unit design).
Expand objectives beyond annihilation (territory, escort, resource denial).
Integrate visualization layer for human spectators.
Allow evolutionary or reinforcement learning meta-strategies.
Multi-simulation tournaments for statistical evaluation.

A Game Design Document for an idea I had.