transfer

Elastic Weight Consolidation

• Preserve weights according to importance to previous task(s)
• Papers:
  • Enabling Continual Learning in Neural Networks

Progressive nets

• Papers:
  • Sim-to-Real Robot Learning from Pixels to Progressive Nets
  • PathNet

Policy Distillation

• Idea: Transferring knowledge of multiple reinforcement learning policies into a single multi-task policy
• Vanilla policy distillation: transfer policies learned from multiple teacher networks to one student network via supervised regression (state, q-value pairs).
• Training student Q-network by matching the output distributions between student/teacher Q-networks using KL-divergence is more effective.
• Hierarchical Experience Replay: Directly combine layers from the teacher networks to the multi-task policy network + new sampling approach for ER
• Papers:
  • Knowledge Transfer for Deep Reinforcement Learning with Hierarchical Experience Replay
  • Actor-Mimic: Deep Multitask and Transfer Reinforcement Learning
  • Policy distillation

Other

Transfer from Simulation to Real World through Learning Deep Inverse Dynamics Model

Learning and Transfer of Modulated Locomotor Controllers

learning to do laps with reinforcement learning and neural nets