Deep Reinforcement Learning optimization with Connection Pruning Techniques

It helps optimizing Deep Reinforcement Learning (DRL) by providing data* on the effect of implementing connection pruning techniques. *It provides example data on how the model behaves. Since DRL tasks are randomized, the behavioral patterns of the model are likely to be similar when implementing pruning techniques compared to when none are used. *added new code 2026 January 2026 Update on A2C Connection Pruning Experiments After discussions with fellow practitioners, it was requested to introduce an alternative method to mimic connection pruning. Instead of relying on increasing dropout rates to simulate temporarily disabling connections during training, the current update implements a soft reset, where the lowest-magnitude weights are explicitly set to zero. This approach provides a more controlled and interpretable form of connection pruning while maintaining network stability. After extensive discussion and review, this update introduces a more explicit and controlled form of connection pruning: Forced Progressive Connection Pruning After each N actions, a fixed fraction of the lowest-magnitude weights in the network are set to zero (effective reset). A safety limit ensures that at least 50% of connections remain in every layer, preventing catastrophic network failure. This implements true, pruning (effective reset) rather than the temporary stochastic masking of dropout. By removing the weakest weights, the network gradually emphasizes the most important connections, reducing redundancy. Progressive pruning applied after each N actions ensures stability while the agent learns. cc: martinkaloev2023@gmail.com