Struct NfspConfig 

pub struct NfspConfig {
    pub max_iterations: usize,
    pub anticipatory_param: f32,
    pub reservoir_capacity: usize,
    pub br_train_steps_per_iteration: usize,
    pub avg_policy_train_steps_per_iteration: usize,
    pub avg_policy_minibatch_size: usize,
    pub avg_policy_lr: f64,
    pub avg_policy_min_reservoir_coverage: f32,
    pub br_reward_scale: f32,
    pub seed: u64,
}

NFSP trainer configuration.

Fields

max_iterations: usize

Number of NFSP outer-loop iterations.

anticipatory_param: f32

η, the anticipatory-mixing parameter. Probability of sampling from the best-response policy (and appending the action to the reservoir) at each rollout step. Heinrich & Silver §3 recommend η = 0.1 for 2-player zero-sum imperfect-information games.

reservoir_capacity: usize

Maximum reservoir buffer size (per agent in the multi-agent setting). Paper default is 2×10⁶; the matching-pennies smoke test runs with ~2000.

br_train_steps_per_iteration: usize

Number of BR train steps (one rollout + one PPO update per step) per outer iteration.

avg_policy_train_steps_per_iteration: usize

Number of supervised cross-entropy steps applied to the AP per outer iteration (skipped when the reservoir is empty).

This is the floor on AP steps. When avg_policy_min_reservoir_coverage is > 0 the trainer may run more than this many steps so the AP sees enough of the reservoir per iteration (issue #199).

avg_policy_minibatch_size: usize

Average-policy supervised minibatch size.

avg_policy_lr: f64

Average-policy learning rate.

avg_policy_min_reservoir_coverage: f32

Minimum reservoir coverage per outer iteration for the AP supervised update, expressed as a multiple of the current reservoir size (issue #199).

NFSP’s average policy is fit by sampling minibatches from a reservoir that grows into the thousands of entries, while the default budget of avg_policy_train_steps_per_iteration × avg_policy_minibatch_size may only touch a few hundred samples per iteration — far too few gradient steps to fit a moving target, so the AP loss stays pinned at the uniform-entropy floor (ln(num_joint_actions)). On bucket-brigade this manifested as avg_ap_loss ≈ ln(40) for an entire 48-iteration run.

When coverage > 0, the trainer runs max(avg_policy_train_steps_per_iteration, ceil(coverage × reservoir_len / avg_policy_minibatch_size)) supervised steps for each agent each iteration, so the AP is exposed to (in expectation) coverage full passes over the reservoir. coverage = 0.0 disables the adaptive floor and preserves the legacy fixed-step behavior. Default 0.0.

br_reward_scale: f32

Optional reward scaling applied to per-step rewards before the BR (PPO) update, to keep the large-magnitude bucket-brigade payoff band ([−700, 0]) from dominating value targets and advantage normalization (issue #199).

Rewards in the anticipatory rollout are multiplied by this factor before being handed to the joint PPO update. 1.0 (default) is a no-op; a value like 0.01 rescales the bucket-brigade band to roughly [−7, 0]. Scaling rewards uniformly does not change the optimal policy (it is an affine transform of the return), but it does keep the critic’s regression targets and the advantage statistics in a numerically friendlier range.

seed: u64

RNG seed (η-flips, reservoir eviction, AP minibatch sampling).

Trait Implementations

impl Clone for NfspConfig

fn clone(&self) -> NfspConfig

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for NfspConfig

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Default for NfspConfig

fn default() -> Self

Returns the “default value” for a type.