Welcome to bnelearn’s documentation! bnelearn is a framework for equilibrium learning in sealed-bid auctions and other markets that can be modeled as Bayesian Games.
This is a work in progress, so the documentation is still incomplete, and may even be outdated or specific to our local deployments in some places. We’ll be working to bring the documentation up to paar in the near future. If you need additional help in the meantime, please get in touch with one of the maintainers.
The public version of the bnelearn repository is hosted at https://github.com/heidekrueger/bnelearn. Development is ongoing via a private GitLab repository on our university server. If you have questions, find an error, or want to contribute, please get in touch via the Issues on the GitHub repository above, or email one of the maintainers.
You can find the installation instructions at Installation and Reproduction Instructions.
A quickstart guide is provided at Quickstart.
Background information can be found under Auction Theory and Equilibrium Computation.
In this section, we will present the bnelearn package with its most essential features and the auction games and learning algorithms it contains.
I(PV) and non-PV (e.g., common values), with arbitrary priors/correlation profiles, utilities, valuation/observation/bid dimensionalities.
Modular organization allows for easy construction of new markets (e.g., bid languages, etc.) from existing or custom building blocks.
Extensive metrics (learning-related: estimates of “equilibrium quality”, utilities over time, market analysis: efficiency, revenue, individual payoffs) and built-in plotting capacities.
Wide range of predefined settings and building blocks:
Learning rules: Policy Gradients, NPGA, PSO.
Auctions: Single-item, multi-unit, LLG combinatorial auction, LLLLGG combinatorial auction.
Priors/correlations: Uniform and normal priors that are either independent or Bernoulli or constant weight dependent.
Utility functions: Quasi-linear utility (risk-neutral) , risk averse, or risk seeking.
Fully vectorized, CUDA enabled, massive parallelism
For combinatorial auctions: Custom batched, CUDA-enabled QP solver for quadratic auction rules and Gurobi/Cvxpy integration for arbitrary auctions stated as a MIP.
Predefined Auction Settings¶
A diverse set of auction games is implemented in the framework.
Algorithms for trying to iteratively learn equilibria implement the base class
Learner in the framework. Two noteworthy algorithms that are contained are
Limitations and Alternatives¶
All players in a game must have the same dimensionality (i.e. same-dimensional type-spaces and action-spaces). Current implementations and learners use deterministic/pure continuous actions.
Other existing multi-agent Learning Packages: Other multi-agent learning frameworks, such as OpenSpiel that is a collection of games and algorithms for reinforcement learning and planning or PettingZoo that is a multi-agent extension of the famous OpenAI Gym framework, mainly focus on zero-sum games and on games with discrete action spaces. Crucially, they neither allow an efficient evaluation of running a large batch of games in parallel.