Scalable agent-based models for optimized policy design: applications to the ec...
summary
Built a transpiler called Salvo that compiles an agent‑based DSL into Jax tensors to run on GPUs/TPUs, reducing a complex ABM implementation from about 9,000 to 2,000 lines and matching or exceeding runtimes of LSD, Agents.jl and FLAME. Calibration and policy search use backprop when agent functions are differentiable and reinforcement learning with Gaussian‑process surrogates when not; a learned flocking policy trains ~50,000 parameters and successfully shifts the flock centroid, demonstrating scalable optimization.
AI-written summary · about
chapters (13)
- 0:00Scalable agent-based models
- 1:06Agent-based models for policy design
- 3:10Gaps in ABM frameworks
- 4:57Criteria for policy-scale ABMs
- 6:20Mathematical formalization of ABMs
- 8:46Differentiable ABMs and optimization
- 9:50Salvo uses JAX for acceleration
- 11:53Implementing complex models in Salvo
- 13:40Circles ABM example in Salvo
- 14:22Salvo vs other frameworks
- 15:28Policy optimization with backpropagation
- 16:51Future research for Salvo
- 17:33Q&A
related talks
37mMultiagent Reinforcement Learning: Rollout and Policy Iteration
46mArchitecting Efficient Agents with Semantic Routing and Mixture-of-Tools
15mSequential Decision Making Problems in Multi-Agent Environments
01h 03mScaling up machine learning without tears (and what do programming languages ha...
42mReasonable AI
30mLearning Multi-Agent Collaborations With Decomposition
28mAgentic AI
15mMulti-Agent Reinforcement Learning In Stochastic Games: From Alphago To Robust Control