Constrained Policy Gradient Method for Safe and Fast Reinforcement Learning: a Neural Tangent Kernel Based Approach

Link: https://arxiv.org/abs/2107.09139

Description: Source code for the NTK based constrained REINFORCE algorithm in the Cartpole OpenAi Gym environment. Constrained_PG_CartPole.py trains the agent defined in PolicyNet.py. Constraint types can be selected with the variable CONSTRAINT_TYPE. Constrained points alongside with some helper functions are included in Functions.py. An example trained agent is added in cartpole_agent_ineq.p that can be tested with TestAgent_CartPole.py.