meta_a2c

Meta-A2C

This is a change on top of the A2C algorithm from the OpenAI baselines as of commit 1b09243. For now, the code is customized to work with the NASGym.

• Original meta-A2C paper: https://arxiv.org/abs/1611.05763
• Original A2C paper: https://arxiv.org/abs/1602.01783
• Example of usage: python -m baselines.run --alg=meta_a2c --env=PongNoFrameskip-v4 --network=meta_lstm [...].

Changes in files with respect to the original A2C

• policies.py: Accepts more inputs for the meta approach
• meta_a2c.py: Instantiates the meta-network and runs adapts the batches of experiments to provide the meta inputs.
• runner.py: Generates the meta inputs.