talks | works
Parabellum

Parabellum

May 6, 2025
Thun, Switzerland
Press D to download as a PDF and F to toggle presentation mode. Navigate up and down with K and J when presenting
PARABELLUMNoah SyrkisMay 6, 20251 |Introduction2 |Vectorization3 |Acceleration4 |Differentiation5 |Simulation6 |Application1 |Introductionβ–ΆSandbox for large-scale, team-based wargames on real terrainβ–ΆDifferentiable JAX environment grounded in OpenStreetMapβ–ΆBuilt for fast iteration on autonomous tactics and analysisβ–ΆA first step towards a digital twin of the battlefield1 of 91 |Introductionβ–ΆTraditional wargaming is not ready for ML, real geography, or beyond real time simulationβ–ΆManual setup slows sensitivity analysisβ–ΆGradient-free simulators block learning-based planning and seamless integration with deep learning pipelines (including LLMs and RL agents)2 of 92 |Vectorizationβ–ΆProcedurally load maps + buildings for any geocoded area in a JAX [1] arrayβ–ΆUnits, teams, sensors as YAML config-files to specify game rules and team compositionsβ–ΆEntirely in JAX: batching, autodiff, vectorisationβ–ΆParallel rollouts across seeds + scenarios on accelerators3 of 93 |Accelerationβ–ΆReal war: high fidelity but slow, costly, unparallelizableβ–ΆParabellum: an RTS-like simulator where:β–ΆArbitrary numbers of sims can run in parallelβ–ΆFaster than real-timeβ–ΆTens of thousands of units per scenario4 of 94 |Differentiationβ–ΆFully written in JAX [1]β–ΆVectorized via vmap, parallelized with pmapβ–ΆDirect integration into deep learning pipelinesβ–ΆBoosts model capacity for long-horizon strategy5 of 95 |Simulationβ–ΆTrajectories as (𝑠𝑑, π‘Žπ‘‘)–tuplesβ–ΆNo rewards Figure 1 β€” only flows of state andactionState 𝑠𝑑+1State 𝑠𝑑Observation π‘œπ‘‘Action π‘Žπ‘‘Step 𝑑Figure 1: Rewardless partially observable MDPdiagram6 of 95 |Simulationβ–ΆState = (position, health, cooldown)β–ΆScene encodes terrain, ranges, unit typesβ–ΆAny Earth location loadable via OSM1β–ΆObservation = visible units’ location, health, type, team1OpenStreetMap data7 of 96 |Applicationβ–ΆHIVE: behavior tree unit controlβ–Άllllll1: foundation-model command & controlβ–ΆNebellum2: tracking rules of engagement1llllll.syrkis.com2nebellum.com8 of 9References[1]J. Bradbury et al., β€œJAX: Composable Transformations of Python+NumPy Programs.” 2018.9 of 9