We study cross-task knowledge reuse in deep reinforcement learning using three complementary paradigms:

• Transfer Learning: pretrain on a source task, then adapt to a target task
• Meta-Learning: learn an initialization that adapts quickly (few-shot) to a task
• Continual Learning: learn tasks sequentially while reducing catastrophic forgetting

Environments used:

• Snake (custom)
• PuckWorld (custom)
• Pong (Atari: ALE/Pong-v5)

Plus a controlled sine-wave regression benchmark (for EWC + toy MAML sanity checks).

python -m venv .venv
source .venv/bin/activate

pip install -r requirements.txt

Gymnasium Atari typically needs ROM install/acceptance.

Recommended:

pip install "gymnasium[atari,accept-rom-license]"

Or AutoROM:

pip install "autorom[accept-rom-license]"
AutoROM --accept-license

Sanity check:

python -c "import gymnasium as gym; env=gym.make('ALE/Pong-v5'); env.reset(); print('Pong OK')"

Train PPO on one environment:

python -m bridging_tasks.transfer.train --env pong --timesteps 1000000
python -m bridging_tasks.transfer.train --env snake --timesteps 500000
python -m bridging_tasks.transfer.train --env puckworld --timesteps 500000

Outputs:

• outputs/transfer/<env>/<timestamp>/
  • saved SB3 model (.zip)
  • reward curve plot
  • tensorboard logs

The original project report frames meta-learning as MAML-style adaptation. In practice, implementing full second-order MAML in RL is expensive and brittle, so this repo includes a first-order meta-learning loop (Reptile-style) that still learns an initialization that adapts quickly.

python -m bridging_tasks.meta.train --iterations 200 --k_shots 5

Outputs:

• outputs/meta/<timestamp>/ (plots + init checkpoints)

Runs:

• sine transfer baselines (scratch / freeze / finetune)
• toy sine MAML
• EWC forgetting matrix

python -m bridging_tasks.continual.run --all

Outputs:

• outputs/continual/<timestamp>/

MIT (See LICENSE).