Reward Assignment: Develop a neural network served as a judge to evaluate the quality of the generated sketch with 83% accuracy.
Action Vector Encoding: Propose an actions vector that can be used by the agent to draw lines and curves.
Model Implementation and Testing: Implement the model using TensorFlow on Python. Modify the code for performance. Use Deep Deterministic Policy Gradients (DDPG) as the reinforcement learning algorithm.
Evaluation: Compare the model with Sketch-RNN as a major benchmark. Over 50% less training steps and more flexible training data compare to Sketch-RNN.