Skip to main content

MetaSolver

thrust_rl::multi_agent::psro

Trait MetaSolver 

Source 
pub trait MetaSolver {
    // Required methods
    fn solve(&self, payoffs: &[Vec<f32>]) -> Vec<f32>;
    fn name(&self) -> &'static str;

    // Provided method
    fn solve_n_player(
        &self,
        _payoffs: &[Vec<f32>],
        num_agents: usize,
        _per_role_k: usize,
    ) -> Vec<f32> { ... }
}
Expand description

Meta-solver over a symmetric 2-player zero-sum empirical game.

Given an n × n row-player payoff matrix payoffs[i][j] representing the expected return of row-player strategy i versus column-player strategy j, returns the row-player’s mixed-Nash distribution as a length-n probability vector summing to 1.0.

For symmetric zero-sum games (matching pennies, the homogeneous-policy version of bucket brigade) the column-player’s equilibrium is the same distribution by symmetry — callers can use the row distribution for both agents. For non-symmetric games, this trait is invoked twice (once per agent role) with appropriately transposed payoff matrices.

Required Methods§

Source

fn solve(&self, payoffs: &[Vec<f32>]) -> Vec<f32>

Solve for the row-player mixed-Nash on a symmetric n × n empirical payoff matrix.

§Contract
  • Input is assumed to be n × n and square; non-square inputs produce undefined behaviour (impl is free to panic).
  • Return vector has length n with non-negative entries summing to 1.0 (within 1e-6 tolerance).
Source

fn name(&self) -> &'static str

Human-readable name for diagnostics / logging.

Provided Methods§

Source

fn solve_n_player( &self, _payoffs: &[Vec<f32>], num_agents: usize, _per_role_k: usize, ) -> Vec<f32>

N-player solve over an explicit per-agent payoff tensor.

payoffs is shape (k^num_agents, num_agents) where payoffs[s][a] is agent a’s payoff at joint pure strategy s. The flat joint-strategy index decomposes into per-agent indices via little-endian mixed-radix (agent 0 = fastest): s = Σ_i s_i · k^i.

Returns a probability vector of length k^num_agents summing to 1.0 ± 1e-6.

§Default

The default implementation only supports num_agents == 2 and delegates to solve via the row-marginal projection. For num_agents > 2 it panics with a message naming the solver. Only AlphaRankMetaSolver overrides this method with a true N-player path; the other in-tree solvers (UniformMetaSolver, FictitiousPlayMetaSolver, ReplicatorDynamicsMetaSolver) have no N>2 generalization with the same convergence guarantees and intentionally panic.

Dyn Compatibility§

This trait is dyn compatible.

In older versions of Rust, dyn compatibility was called "object safety".

Implementors§