ahmedo42 / autockt Goto Github PK

The idea is simple ,a script that takes a bunch of specs , loads a trained model and runs the algorithm to get the final parameters

In order to intuitively understand the generated specs and the actions taken by the agent we need to be able to visualize what's happening in the environment.

• Visualize Distribution of Generated Specs ( validation + training )

• Visualize Distribution of Actions for each circuit parameter

Hi, @Ahmed, do you try using this Autockt frame to optimize more than 6 circuit specs? I did it but I found that my reward dropped from -170 to -50 (it took more than three days), it basically no longer converged. It's been running for 8 days now and it's still floating around -50. The program terminate condition is "stop": {"episode_reward_mean": -0.02}. Some settings were shown as below:
"num_workers" : 0, "horizon" : 50, "batch_size" : 1200, "framework" : "torch", "algo" : "PPO", "episodes_per_batch" : 50,

The idea is to compare BagNet ( Genetic + NN) to AutoCkt (RL).

BagNet Source Code

To make sure that the results are authentic or not