Keras Reinforcement Learning Projects

Hindsight Experience Replay

One of the abilities humans have is to learn from our mistakes and adapt next time to avoid making the same mistake. This is the basis of reinforcement-learning algorithms. The greatest difficulties in implementing these algorithms are encountered when dealing with scattered prizes. Consider the following scenario: a learning agent must control a robot arm to open a box and place an object inside it. While defining the reward for this task is simple and straightforward, the underlying learning problem is difficult. The agent must uncover a long sequence of correct actions to find a configuration of the environment that produces the sparse reward—the object placed inside the box. Discovering this poor reward signal ...