arushir

# Run in command line:
# mindmeld blueprint hr_assistant
# python -m hr_assistant build


# MindMeld NLP tutorial
bp_path = 'hr_assistant'
from mindmeld.components.nlp import NaturalLanguageProcessor
nlp = NaturalLanguageProcessor(app_path=bp_path)
nlp.load()

arushir

I implemented the DQN model from this paper: https://www.cs.toronto.edu/~vmnih/docs/dqn.pdf.

I used a simple network with two hidden layers and an output layer, instead of the CNN described in the paper due to the relative simplicity of the Cart-Pole environment compared to Atari games.

I implemented the target network described in the more recent paper here: https://storage.googleapis.com/deepmind-data/assets/papers/DeepMindNature14236Paper.pdf.

Note that compared to the implementation without the target network found here https://gym.openai.com/evaluations/eval_EdDdHKTo6OsW4v0MukA, this model takes longer to train but the training is smoother and it is more stable once the problem is solved.

arushir

I implemented the DQN model from this paper: https://www.cs.toronto.edu/~vmnih/docs/dqn.pdf.

I used a simple network with two hidden layers and an output layer, instead of the CNN described in the paper due to the relative simplicity of the Cart-Pole environment compared to Atari games.

Note, that I did not yet implement the target network described in the more recent paper here: https://storage.googleapis.com/deepmind-data/assets/papers/DeepMindNature14236Paper.pdf.

The results vary from run to run, sometimes taking 1000 episodes to solve the problem, and at other times taking only 200 episodes.