talks
|
works
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
PARABELLUM
Noah Syrkis
May 6, 2025
1 |
Introduction
2 |
Vectorization
3 |
Acceleration
4 |
Differentiation
5 |
Simulation
These slides presents a broad overview of the Parabellum environment shown in
[1] T. Anne
et
al.
, βHarnessing Language for Coordination: A Framework and Benchmark for LLM-Driven Multi-
Agent Control,β
IEEE Transactions on Games
, pp. 1β25, 2025, doi:
10.1109/TG.2025.3564042
.
This
research is funded by Armasuisse.
1 of
9
1 |
Introduction
βΆ
Sandbox for large-scale, team-based wargames on real terrain
βΆ
Differentiable JAX environment grounded in OpenStreetMap
βΆ
Built for fast iteration on autonomous tactics and analysis
2 of
9
1 |
Introduction
βΆ
Traditional wargaming is not ready for ML, real geography, or beyond real time simulation
βΆ
Manual setup slows sensitivity analysis
βΆ
Gradient-free simulators block learning-based planning and seamless integration with deep learn
ing pipelines (including LLMs and RL agents)
3 of
9
2 |
Vectorization
βΆ
Procedurally load maps + buildings for any geocoded area in a JAX
[2]
array
βΆ
Units, teams, sensors as YAML config-files to specify game rules and team compositions
βΆ
Entirely in JAX: batching, autodiff, vectorisation
βΆ
Parallel rollouts across seeds + scenarios on accelerators
4 of
9
3 |
Acceleration
βΆ
Real war: high fidelity but slow, costly, unparallelizable
βΆ
Parabellum: an RTS-like simulator where:
βΆ
Arbitrary numbers of sims can run in parallel
βΆ
Faster than real-time
βΆ
Tens of thousands of units per scenario
5 of
9
4 |
Differentiation
βΆ
Fully written in JAX
[2]
βΆ
Vectorized via
vmap
, parallelized with
pmap
βΆ
Direct integration into deep learning pipelines
βΆ
Boosts model capacity for long-horizon strategy
6 of
9
5 |
Simulation
βΆ
Trajectories as (
π
π‘
,
π
π‘
)βtuples
βΆ
No rewards
Figure 1
β only flows of state and
action
State
π
π‘
+
1
State
π
π‘
Observation
π
π‘
Action
π
π‘
Step
π‘
Figure 1: Rewardless partially observable MDP
diagram
7 of
9
5 |
Simulation
βΆ
State = (position, health, cooldown)
βΆ
Scene encodes terrain, ranges, unit types
βΆ
Any Earth location loadable via OSM
1
βΆ
Observation = visible unitsβ location, health, type, team
1
OpenStreetMap data
8 of
9
References
[1]
T. Anne
et al.
, βHarnessing Language for Coordination: A Framework and Benchmark for
LLM-Driven Multi-Agent Control,β
IEEE Transactions on Games
, pp. 1β25, 2025, doi:
10.1109/
TG.2025.3564042
.
[2]
J. Bradbury
et al.
, βJAX: Composable Transformations of Python+NumPy Programs.β 2018.
9 of
9
PARABELLUM
Noah Syrkis
May 6, 2025
1 |
Introduction
2 |
Vectorization
3 |
Acceleration
4 |
Differentiation
5 |
Simulation
These slides presents a broad overview of the Parabellum environment shown in
[1] T. Anne
et
al.
, βHarnessing Language for Coordination: A Framework and Benchmark for LLM-Driven Multi-
Agent Control,β
IEEE Transactions on Games
, pp. 1β25, 2025, doi:
10.1109/TG.2025.3564042
.
This
research is funded by Armasuisse.
1 of
9
1 |
Introduction
βΆ
Sandbox for large-scale, team-based wargames on real terrain
βΆ
Differentiable JAX environment grounded in OpenStreetMap
βΆ
Built for fast iteration on autonomous tactics and analysis
2 of
9
1 |
Introduction
βΆ
Traditional wargaming is not ready for ML, real geography, or beyond real time simulation
βΆ
Manual setup slows sensitivity analysis
βΆ
Gradient-free simulators block learning-based planning and seamless integration with deep learn
ing pipelines (including LLMs and RL agents)
3 of
9
2 |
Vectorization
βΆ
Procedurally load maps + buildings for any geocoded area in a JAX
[2]
array
βΆ
Units, teams, sensors as YAML config-files to specify game rules and team compositions
βΆ
Entirely in JAX: batching, autodiff, vectorisation
βΆ
Parallel rollouts across seeds + scenarios on accelerators
4 of
9
3 |
Acceleration
βΆ
Real war: high fidelity but slow, costly, unparallelizable
βΆ
Parabellum: an RTS-like simulator where:
βΆ
Arbitrary numbers of sims can run in parallel
βΆ
Faster than real-time
βΆ
Tens of thousands of units per scenario
5 of
9
4 |
Differentiation
βΆ
Fully written in JAX
[2]
βΆ
Vectorized via
vmap
, parallelized with
pmap
βΆ
Direct integration into deep learning pipelines
βΆ
Boosts model capacity for long-horizon strategy
6 of
9
5 |
Simulation
βΆ
Trajectories as (
π
π‘
,
π
π‘
)βtuples
βΆ
No rewards
Figure 1
β only flows of state and
action
State
π
π‘
+
1
State
π
π‘
Observation
π
π‘
Action
π
π‘
Step
π‘
Figure 1: Rewardless partially observable MDP
diagram
7 of
9
5 |
Simulation
βΆ
State = (position, health, cooldown)
βΆ
Scene encodes terrain, ranges, unit types
βΆ
Any Earth location loadable via OSM
1
βΆ
Observation = visible unitsβ location, health, type, team
1
OpenStreetMap data
8 of
9
References
[1]
T. Anne
et al.
, βHarnessing Language for Coordination: A Framework and Benchmark for
LLM-Driven Multi-Agent Control,β
IEEE Transactions on Games
, pp. 1β25, 2025, doi:
10.1109/
TG.2025.3564042
.
[2]
J. Bradbury
et al.
, βJAX: Composable Transformations of Python+NumPy Programs.β 2018.
9 of
9
