This is very naive implementation of Learning to Learn without Gradient Descent by Gradient Descent with A3C.

5000 tasks(scale0.1_wh50_5000.p, scalerand_wh50_5000.p) are instantiated from gaussian process(2dim.) with fixed scale(0.1) and random scale([0.01,0.61)).

These tasks are 50 X 50.

Each instance(task) looks like

4900 episodes for training / 100 episodes for test

• Agent

  • is 2 layer LSTM.
  • has continuous output (beta distribution).
  • gets output & reward of previous step as input.

• Each agent chooses one task from 4900 training tasks randomly at the start of episode

• Similarly, each agent chooses one task from 100 test tasks randomly at test episode.

• tensorflow==1.13.2
• scikit-learn==0.21.3

python main_a3c_gp_meta.py