Presents MARL environments for realistic power-market bidding that incorporate grid-constrained clearing and GPU-accelerated solvers.
PowerMarketJax is an open-source benchmark suite for Multi-Agent Reinforcement Learning (MARL) in power markets, implemented entirely in JAX to enable GPU-accelerated training and evaluation. It features five distinct market environments—day-ahead wholesale, real-time balancing, ancillary services, peer-to-peer double auctions, and local flexibility—each preserving specific clearing, pricing, and settlement rules under a unified interface.
By integrating optimization-based market clearing and policy training into a single compiled computation graph, PowerMarketJax eliminates the need for CPU-based solvers and reduces host-device synchronization overhead. This architecture achieves up to 33x speedup over CPU-based baselines, allowing for large-scale parallelism across 1,024 environments with 1,200 participants simultaneously. The suite includes independent learners like PPO and SAC, providing a standardized framework to study how market design influences strategic bidding behavior under grid physical constraints.
PowerMarketJax introduces a JAX-based benchmark suite for multi-agent reinforcement learning in electricity-market settings. Rather than treating bidding as an abstract auction problem, the environments couple strategic market participation with grid-constrained clearing: agents submit bids or offers, and a market-clearing layer determines dispatch and pricing while respecting network constraints. The use of JAX and GPU-accelerated solvers is central to the design, enabling vectorized simulation, high-throughput rollouts, and more realistic market physics than simplified MARL testbeds.
The key contribution is a standardized, computationally efficient platform for studying MARL in power markets. The suite makes explicit the intersection of learning, market design, and power-system optimization by requiring agents to operate in a setting where network congestion, clearing feasibility, and other agents’ bids jointly shape rewards. This setup exposes several practical challenges for MARL, including nonstationarity from strategic interaction, partial observability, constrained decision-making, and the computational cost of realistic market-clearing. It also highlights an important insight: bidding policies that perform well in unconstrained or toy market environments may behave very differently once grid constraints materially affect clearing outcomes.
The work matters because it bridges two areas that have often been studied separately: reinforcement learning for strategic decision-making and optimization-based power-market operation. By providing a reproducible JAX/GPU benchmark, it lowers the barrier to training and comparing MARL algorithms under realistic market conditions. This could support research on learning-based bidding, price formation, congestion management, and market efficiency, and more broadly demonstrate how modern accelerated simulation stacks can make domain-specific MARL benchmarks both realistic and scalable.