Struct JointRollout 

pub struct JointRollout {
    pub observations_per_agent: Vec<Vec<f32>>,
    pub obs_dims: Vec<usize>,
    pub actions: Vec<Vec<i64>>,
    pub num_action_dims: usize,
    pub log_probs: Vec<Vec<f32>>,
    pub values: Vec<Vec<f32>>,
    pub rewards: Vec<Vec<f32>>,
    pub dones: Vec<f32>,
}

Synchronized rollout buffer (host-side).

Per-agent observations are stored independently so environments with distinct per-agent views (e.g. partial observability, asymmetric information) work without any pre-concatenation. Each agent i records its own observation stream into observations_per_agent[i] (a flat [T * obs_dim] buffer). Per-agent actions / log-probs / values / rewards are stored as parallel host buffers and materialized into Burn tensors lazily inside JointMultiAgentTrainer::update.

Fields

observations_per_agent: Vec<Vec<f32>>

Per-agent observations: observations_per_agent[i] is a flat [T * obs_dims[i]] buffer holding the observation stream for one agent across the rollout.

obs_dims: Vec<usize>

Per-agent observation dimensionality (obs_dims[i] = length of agent i’s observation vector). Heterogeneous across agents is supported — e.g. a SignalingGame speaker observes a length-V one-hot while the listener observes a single received-token slot (issue #275).

actions: Vec<Vec<i64>>

Per-agent actions: Vec<N>[T * num_action_dims]. num_action_dims is 1 for scalar discrete, num_dims for multi-discrete.

num_action_dims: usize

Number of action dimensions (uniform across agents in this first cut).

log_probs: Vec<Vec<f32>>

Per-agent rollout-time log-probs: Vec<N>[T].

values: Vec<Vec<f32>>

Per-agent value estimates: Vec<N>[T].

rewards: Vec<Vec<f32>>

Per-agent rewards: Vec<N>[T].

dones: Vec<f32>

Episode-termination flag (shared across agents): [T].

Implementations

impl JointRollout

pub fn num_steps(&self) -> usize

Rollout length in steps.

pub fn num_agents(&self) -> usize

Number of agents represented in this rollout.

Trait Implementations

impl Clone for JointRollout

fn clone(&self) -> JointRollout

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for JointRollout

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

Formats the value using the given formatter.