Struct JointStats 

pub struct JointStats {
    pub policy_loss: Vec<f64>,
    pub value_loss: Vec<f64>,
    pub entropy: Vec<f64>,
    pub clip_fraction: Vec<f64>,
    pub approx_kl: Vec<f64>,
    pub explained_var: Vec<f64>,
    pub aux_loss: f64,
    pub total_loss: f64,
    pub num_mb_steps: usize,
}

Per-update training statistics for the joint trainer.

Mirrors crate::train::ppo::TrainingStats but with per-agent vectors for the agent-local quantities and a single shared scalar for the auxiliary cross-agent term.

Fields

policy_loss: Vec<f64>

Per-agent policy loss (averaged across PPO epochs).

value_loss: Vec<f64>

Per-agent value-function loss.

entropy: Vec<f64>

Per-agent entropy.

clip_fraction: Vec<f64>

Per-agent fraction of clipped updates.

approx_kl: Vec<f64>

Per-agent approximate KL divergence between old and new policy.

explained_var: Vec<f64>

Per-agent explained variance of the value function.

aux_loss: f64

Auxiliary cross-agent loss (e.g. λ * redundancy_penalty). Scalar shared by all agents because it’s computed jointly on the same minibatch features.

total_loss: f64

Total summed loss Σ_i agent_loss_i + aux_loss (averaged across PPO epochs).

num_mb_steps: usize

Number of minibatch gradient steps actually taken across the whole update (Σ_epochs minibatches walked). For the default single-minibatch path this equals n_epochs; with JointTrainerConfig::iterate_all_minibatches it equals n_epochs × ceil(num_steps / minibatch_size), and with JointTrainerConfig::max_minibatches_per_epoch = Some(cap) it is reduced to n_epochs × min(cap, num_minibatches). Exposed so the throughput lever (issue #251) is observable/testable.

Implementations

impl JointStats

pub fn zeros(num_agents: usize) -> Self

Construct a fully-zeroed JointStats sized for num_agents agents.

Trait Implementations

impl Clone for JointStats

fn clone(&self) -> JointStats

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for JointStats

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

Formats the value using the given formatter.

impl Default for JointStats

fn default() -> JointStats

Returns the “default value” for a type.