Module reward_shaping 

Potential-based reward shaping (PBRS) and goal-distance potentials.

Implements the PBRS transform r' = r + gamma * Phi(s') - Phi(s) which preserves the optimal policy (Ng et al., 1999), plus goal-distance potential computation for goal-conditioned RL.

Structs

GoalDistanceTransform

Goal-distance reward transform.

PBRSTransform

Potential-based reward shaping transform.

RewardContext

Context passed to reward transforms for access to transition metadata.

Traits

RewardTransform

Trait for composable reward transformations.

Functions

compute_goal_distance_potentials

Goal-distance potential: Phi(s) = -scale * ||s[goal_slice] - goal||_2

shape_rewards_pbrs

Compute shaped rewards: r' = r + gamma * Phi(s') - Phi(s)