PCPPO explores using reinforcement-learning agents to choose structural changes in a predictive-coding network. The predictive-coding dynamics form the environment, while PPO agents select actions intended to alter connections.
The training framework includes a shared actor, a centralized critic, and a vectorized predictive-coding core. The proposed role of the agents is to decide where structural changes would help the network's task.
The structural-action step is still a placeholder in the vectorized implementation. Selected actions do not yet change its dense layers, and pruning is treated as a no-op. This is therefore a framework for investigating structural learning, not a demonstrated self-restructuring network. Applying those actions meaningfully is the next implementation step.