Stochastic Warfare — Brainstorm & Domain Decomposition¶

Simulation Scales (Layered Architecture)¶

1. Campaign Level (Strategic)¶

Theater-wide resource allocation, supply lines, reinforcement scheduling
Attrition models over time (Lanchester models as a starting point?)
Strategic decision trees / game-theoretic opponent modeling
Terrain and geography at macro scale (regions, fronts, LOCs)

2. Battlefield Level (Operational)¶

Force disposition and maneuver across a battlefield
Combined arms interactions
Objective control, flanking, envelopment
Fog of war / information asymmetry modeling

3. Battle Level (Tactical)¶

Engagement resolution between formations
Fire-and-maneuver sequencing
Morale, suppression, cohesion effects
Cover, concealment, line-of-sight

4. Unit Level (Individual/Squad)¶

Individual or small-group behavior
Weapon accuracy, rate of fire, ammunition
Fatigue, injury, skill levels
Micro-terrain interaction

Stochastic & Signal Processing Models (Core Engine Ideas)¶

These are the mathematical tools available for modeling. Each is annotated with its implementation home in specs/project-structure.md.

Random Process Models¶

Markov Chains: State transitions for unit morale, readiness, engagement status → morale/state.py
Poisson Processes: Event arrivals (reinforcements, supply deliveries, random encounters) → simulation/campaign.py, core/events.py
Gaussian/Shot Noise: Weapon accuracy, dispersion patterns, sensor noise → combat/ballistics.py, combat/indirect_fire.py, detection/sensors.py
Brownian Motion / Random Walks: Unit drift in movement under uncertainty, stress accumulation → movement/engine.py, morale/stress.py
Queueing Theory: Supply chain bottlenecks, medical evacuation, repair depots → logistics/medical.py, logistics/maintenance.py, logistics/supply_network.py

Signal Processing Analogies¶

Kalman Filtering: Intelligence estimation — tracking enemy positions with noisy observations → detection/estimation.py
Monte Carlo Methods: Outcome probability estimation for engagements, campaign-level analysis → validation/monte_carlo.py, Phase 7 (engagement) / Phase 10 (campaign) validation
Spectral Analysis: Analyzing periodicity in operational tempo → tools/tempo_analysis.py (post-run analysis tool, not real-time engine component)
Convolution: Modeling cascading effects (suppression spreading through a formation) → combat/suppression.py, morale/cohesion.py
Matched Filtering / Detection Theory: Reconnaissance and detection modeling (SNR-based detection probability) → detection/detection.py

Optimization¶

Linear/Nonlinear Programming: Optimal resource allocation (ammo, fuel, troops) → c2/ai/assessment.py (combat power), logistics/supply_network.py (flow optimization)
Stochastic Optimization: Decision-making under uncertainty → c2/ai/decisions.py
Graph Theory / Network Flow: Supply line optimization, movement planning → logistics/supply_network.py, movement/pathfinding.py
Game Theory: Adversarial decision modeling → c2/ai/decisions.py, c2/ai/stratagems.py (opponent modeling in COA analysis)

Attrition & Force Models¶

Lanchester Models (square law, linear law): Attrition rate baselines for campaign-level modeling, validation benchmarks → simulation/campaign.py (aggregate force modeling), Phase 10 backtesting comparisons

Key Domain Systems to Model¶

Combat Resolution¶

Probability-of-hit models (range, weapon, skill, conditions)
Damage / lethality models
Suppression and morale effects
Armor penetration / protection
Long-range fires & deep strike: tube artillery, rocket artillery (MLRS/HIMARS — area fires and precision-guided munitions), mortars
Surface-to-surface missiles: theater ballistic missiles (Scud, Iskander, DF-21), land-attack cruise missiles (Tomahawk, Kalibr, JASSM), ground-launched anti-ship missiles (coastal defense — Bastion, NSM, shore-based Harpoon)
Missile defense: ballistic missile defense (Patriot PAC-3, THAAD, S-400 BMD mode, Aegis BMD), cruise missile defense, counter-rocket/artillery/mortar (C-RAM, Iron Dome)
Sensor-to-shooter kill chain: targeting cycle from detection through engagement — time-sensitive targeting, dynamic targeting, BDA feedback loop

Movement & Maneuver¶

Terrain-dependent movement rates
Formation movement with stochastic deviation
Pathfinding under uncertainty (fog of war)
Fatigue accumulation

Logistics & Supply¶

Supply chain as a network flow problem
Consumption rates (ammo, fuel, food, medical)
Stockpile management and forward staging
Disruption modeling (interdiction, weather)

Command & Control (C2)¶

Order propagation delays
Communication reliability (noise model)
Decision-making AI for autonomous units
Chain of command and delegation

Intelligence & Reconnaissance¶

Detection as a signal-in-noise problem
Information decay over time
Fog of war implementation
Deception and counter-intelligence

Terrain & Environment¶

Hex grid vs continuous space vs hybrid
Elevation, vegetation, urban, water
Weather effects on movement/combat/visibility
Day/night cycle
Urban/suburban/rural spectrum: density, building types, infrastructure quality
Infrastructure networks: roads, bridges, rail, utilities — function as terrain features AND logistical assets
Civilian population: density and disposition (friendly/neutral/hostile) — affects ROE, logistics, intelligence, morale

Morale & Human Factors¶

Unit cohesion as a state variable
Stress/fatigue accumulation (random walk with drift)
Rout/rally mechanics
Leadership influence radius

Military Theorist Foundations¶

Purpose¶

Ground modeling decisions in established military theory — not just engineering intuition
Validate that simulation dynamics capture the right phenomena
Future feature potential: AI commanders operating according to doctrinal schools

Key Thinkers & Their Relevance to the Engine¶

Thinker	Key Concepts	Engine Relevance
Sun Tzu	Deception, intelligence, terrain taxonomy, morale	Intel/recon system, terrain classification, morale model
Clausewitz	Friction, fog of war, center of gravity, culminating point	Stochastic deviation models, fog of war mechanic, campaign-level objectives
Jomini	Interior/exterior lines, concentration of force, LOCs	Movement/maneuver, logistics network modeling
Lanchester	Mathematical combat models (square/linear laws)	Combat resolution baseline, attrition modeling
Mahan	Sea power, naval logistics, chokepoints, command of the sea	Naval warfare, SLOC control, maritime logistics, chokepoint modeling
Corbett	Maritime strategy, fleet-in-being, sea control vs sea denial	Naval campaign modeling, maritime strategy AI, blockade mechanics
Gorshkov	Soviet naval strategy, balanced fleet, sea denial	Asymmetric naval strategy, submarine warfare doctrine
Wayne Hughes	Fleet tactics, salvo model (missile exchange)	Naval surface combat resolution, anti-ship missile modeling
Douhet	Strategic air power, command of the air	Air superiority modeling, strategic bombing
Liddell Hart	Indirect approach, expanding torrent, maneuver warfare	AI maneuver doctrine, operational planning
Fuller	Mechanized warfare, combined arms	Combined arms interaction rules
Boyd	OODA loop	C2 cycle modeling: sensor → processing → decision → order propagation
Warden	Five rings (strategic targeting)	Target prioritization for strategic/air campaigns
Tukhachevsky	Deep operations, operational depth	Deep fires targeting, operational-depth strike planning
Starry / DePuy	AirLand Battle, deep attack, synchronization	Deep fires doctrine, sensor-to-shooter integration, fire support coordination

Philosophers, Historians & Political Theorists¶

Thinker	Key Concepts	Engine Relevance
Thucydides	Realism, state behavior, escalation dynamics	Campaign-level decision modeling, escalation mechanics
Machiavelli	Political-military nexus, fortune vs preparation	Strategic AI decision-making, contingency planning
Grotius/Vattel	Laws of war, just war theory, proportionality	ROE system, engagement constraints, civilian protection
Walzer	Just and unjust wars, civilian distinction, moral constraints	ROE modeling, civilian casualty tracking, ethical constraints on targeting
Kant	Ethical constraints on state violence	Framework for modeling political constraints on military operations
Hobbes	Security dilemma, rationale for organized force	Strategic-level modeling of conflict initiation and escalation

Doctrinal AI (Future Feature)¶

AI commanders operating according to specific doctrinal schools
e.g. Clausewitzian AI (seek center of gravity) vs Sun Tzu AI (deception + intel advantage) vs maneuverist AI (tempo + indirect approach)
Enables comparative analysis of doctrinal effectiveness in simulated scenarios

Technical Architecture Questions (To Discuss)¶

Event-driven vs tick-based vs hybrid simulation loop?
Spatial representation: hex grid, continuous 2D, or layered?
Data-driven unit definitions (YAML/JSON) vs code?
How to handle multi-scale interaction (strategic events triggering tactical battles)?
Serialization / save-state for long campaigns?
Headless simulation first, UI later?
What historical era(s) to target first for prototyping?

Architecture Decisions¶

1. Simulation Loop — HYBRID¶

Tick-based outer loop at variable resolution depending on the scale being analyzed
Strategic/campaign layer: coarser ticks (hours/days)
Tactical/battle layer: finer ticks (seconds/minutes)
User controls the analysis scale, and tick resolution scales accordingly
Event-driven resolution within ticks for discrete interactions (engagements, arrivals, detections)
Rationale: Purely event-driven is too coarse even at strategic scale — logistics movements, convoy tracking, and supply chain state need continuous tracking (a convoy interdicted mid-route vs. one that arrives are fundamentally different outcomes). Hybrid lets us scale resolution up and down based on what the user is analyzing while maintaining fidelity where it matters.
Tick-based backbone ensures nothing "falls through the cracks" between events

2. Spatial Representation — LAYERED HYBRID¶

Strategic (graph-based): Regions as nodes, routes/corridors as edges. Nodes carry aggregate terrain properties (mountainous, forested, road quality). Natural fit for network flow logistics and strategic movement.
Operational/Tactical (gridded — hex or raster): Cells reference heightmap and terrain classification layers. Supports LOS, fire arcs, movement cost calculations.
Unit level (continuous): Real (x, y, z) coordinates. Interpolated elevation, vector micro-terrain features (buildings, walls, treelines).
Aerial units: 2.5D model — 2D heightmap + altitude as a unit property. Aircraft care about altitude-relative-to-ground (terrain masking, SAM envelopes), not full 3D voxels. Keeps it efficient.
Transitions: User zooms into a region/hex to resolve at finer granularity; the system loads the appropriate spatial model for that scale.

2a. Coordinate System — ENU / UTM (not geodetic)¶

Internal simulation math runs in a local Cartesian frame (ENU or UTM-projected), not geodetic (lat/lon)
Distances in meters, Euclidean geometry, linear algebra — no spherical math overhead
Geodetic (WGS84 lat/lon) used only for: scenario definition, real-world data import, map visualization/export
Projection workflow: scenario defines geographic origin → terrain & positions projected to local Cartesian at load → all sim math in Cartesian → convert to geodetic only for display
UTM as baseline projection: military standard, maps to MGRS, handles operational/campaign scale well. pyproj for conversions.
Local ENU refinement: for tactical/unit-level precision within a UTM zone
Campaign-scale caveat: continental theaters may tile multiple UTM zones or local frames with known transforms. Single-origin distortion is ~0.004% at 500km — within simulation noise for most purposes.
2.5D: altitude stored as a unit property (for aerial) and as heightmap values (for terrain), both in meters above reference

2b. Terrain Data Model¶

Elevation: Heightmaps / Digital Elevation Models (DEMs) — 2D arrays where each cell stores elevation. Industry standard in GIS and military sim. Freely available real-world data (SRTM, ASTER) at ~30m resolution. numpy native.
Classification layers: Stacked attribute channels per cell — land cover (forest, urban, open, marsh, water), road networks, concealment, trafficability. Multi-channel approach, like image data.
Micro-terrain: Vector features (shapely polygons) for building footprints, walls, treelines at unit-level scale.
Derived products: Slope, aspect, LOS, watersheds computed from elevation math. Military terrain analysis (OCOKA) maps naturally onto this.
Multi-scale terrain: Strategic nodes carry aggregate terrain summaries; tactical grid cells carry full heightmap + classification; unit-level interpolates heightmap and uses vector features.
Libraries: numpy (heightmap math, LOS raycasting), scipy.ndimage (terrain analysis), rasterio (real-world GIS import), shapely (vector micro-terrain)

7. Prototype Era — MODERN (Cold War – Present)¶

Modern era exercises all subsystems: combined arms, air power, radar/sensors, guided weapons, EW, complex logistics
Best documentation and data availability for backtesting and validation
Building for modern complexity first means scaling back to earlier eras (WW2, Napoleonic, etc.) is a matter of removing/simplifying layers rather than adding them
Earlier eras remain targets for future support and historical campaign validation

6. Development Approach — HEADLESS ENGINE FIRST¶

Pure Python simulation engine with no UI dependency
Basic Python visualization (matplotlib, simple plots) for validation and debugging during development
Full UI is a separate future effort in another language/framework
Engine exposes clean APIs that a UI layer can consume later
This keeps the core testable, portable, and focused

5. Serialization & Save-State — CHECKPOINT + DETERMINISTIC REPLAY¶

Deterministic simulation: fixed PRNG seed + initial state + user inputs = perfectly reproducible run
No need for per-tick snapshots: replay from nearest checkpoint using saved PRNG state
Periodic checkpoints at natural boundaries: campaign tick boundaries, before/after tactical engagements, user-triggered saves
Each checkpoint = full state snapshot + PRNG state at that point
To inspect any moment: replay forward from nearest prior checkpoint (fast, small window)
PRNG discipline requirements:
All randomness through seeded numpy.random.Generator, never random.random() or system entropy
Dedicated PRNG streams per subsystem (combat, movement, intel, morale) forked from master seed — prevents cross-subsystem sequence contamination
Deterministic iteration order — no unordered sets/dicts driving sim logic
Single-threaded simulation core (parallelism only for independent sub-sims with own PRNG forks)
No timing-dependent behavior in the deterministic path
Serialization: numpy.random.Generator.bit_generator.state captures exact PRNG state as a dict. pickle/msgpack for fast state serialization; h5py/zarr if state grows large.
User input log: all external decisions (orders, setting changes) timestamped and stored alongside checkpoints for full replay fidelity

4. Multi-scale Interaction — FULL TACTICAL RESOLUTION ALWAYS¶

Every engagement runs through the full tactical simulation layer, regardless of what scale the user is analyzing
Fidelity is never sacrificed — the simulation doesn't skip steps or hand-wave outcomes
Visibility vs fidelity: if the user is focused on strategic/logistical analysis, tactical battles still run under the hood; the user sees aggregated results. All detailed data is retained and available for drill-down.
Auto-resolve as optional performance mode: user can toggle for fast campaign previews, with the understanding that accuracy is traded for speed. Not the default.
Interface contract: campaign layer provides inputs (forces, terrain, objectives, intel state) → tactical layer runs full resolution → returns outputs (casualties, ammo/fuel consumed, territory, time elapsed, unit states) → campaign layer integrates results
Campaign tick may spawn tactical sub-simulations, run them, collect results, resume
A priori intelligence: before contact, each side maintains an estimated belief state of the enemy (positions, strength, composition) based on available intel sources (radar, recon, SIGINT, satellites, doctrinal assumptions). This belief state drives pre-contact decisions and may be inaccurate — fog of war as a real mechanic, not a visibility toggle.
Connects to Kalman filter model: belief state updated by noisy observations from sensor/intel assets

3. Unit Definitions — DATA-DRIVEN (YAML + Pydantic)¶

Engine defines broad unit classes (armor, infantry, rotary-wing, fixed-wing, artillery, etc.) that encode behaviors and interaction rules in Python
YAML config files parameterize specific unit types within those classes (M1A2 Abrams is an instance of the armor class with specific speed, protection, weapons, etc.)
Adding new unit types or eras = adding YAML files, no code changes
Pydantic models validate YAML at load time — enforces required fields, value ranges, type correctness
Scenario packs = a folder of unit definition YAMLs + map data, portable and version-controllable
Unit class hierarchy and detailed field specs to be designed when implementation begins
User has additional unit-side ideas to incorporate at implementation time

2c. Aerial Units — Scope Confirmation¶

Aerial units are in scope and essential across all simulation layers
Strategic: airlift, strategic bombing, air superiority campaigns
Operational: CAS, interdiction, battlefield air recon
Tactical: close air support integration, helicopter operations
Unit: individual aircraft behavior, weapon delivery, evasion
Air defense networks modeled as detection/engagement envelopes in 2.5D

Potential Python Libraries¶

pyproj — coordinate system transforms (geodetic ↔ UTM ↔ ENU)
numpy / scipy — core math, distributions, signal processing
networkx — graph/network modeling (supply lines, C2)
simpy — discrete event simulation
shapely / geopandas — spatial/terrain modeling (if continuous)
pygame / pyglet — simple 2D visualization (early prototyping)
pydantic — data validation for unit/scenario definitions
h5py / zarr — efficient state serialization for large simulations