Deep Q network
Budget: $30 – $250 USD
The aim of this project is to develop a reinforcement learning agent that aims at the target container and pours all the objects into it without spillage. You are provided a scene with two cups and two cubes.
If you run the code, the cup should start rotating at a velocity randomly chosen from a predefined set (thee velocities are more diverse than in part two). Use the provided functions to move the pouring cup horizontally while rotating it.
Action Space:
The source cup can be moved along the X-axis by selecting an action from a predefined set [-2, -1,0, 1, 2]. You don’t have to move the cup in any other axis.
State Space:
The state space involves:
The position of two cubes
The position of source cup
Velocities of the two cubes(optional)
Task:
the Q-Learning algorithm to develop the Q-table. For most problems, it is impractical to represent the Q-function as a table containing values for each combination of state and action. Because of that, in this part, you have to train a Deep Q-Network to estimate the Q-values.
Submission:
Submit your training and testing code
Submit a video that shows cubes are falling to the target after you run your code
Submit a .txt file that contain the log of the training, including the accumulated rewards for each episode, TD errors for each DQN update
Submit a .txt file with the following result:
How many times all the cubes fell successfully to the target in 100 trials if your state space consists only cup position?
How many times all the cubes fell successfully to the target in 100 trials if your state space consists both cup position and two cube’s position?
If you run the code, the cup should start rotating at a velocity randomly chosen from a predefined set (thee velocities are more diverse than in part two). Use the provided functions to move the pouring cup horizontally while rotating it.
Action Space:
The source cup can be moved along the X-axis by selecting an action from a predefined set [-2, -1,0, 1, 2]. You don’t have to move the cup in any other axis.
State Space:
The state space involves:
The position of two cubes
The position of source cup
Velocities of the two cubes(optional)
Task:
the Q-Learning algorithm to develop the Q-table. For most problems, it is impractical to represent the Q-function as a table containing values for each combination of state and action. Because of that, in this part, you have to train a Deep Q-Network to estimate the Q-values.
Submission:
Submit your training and testing code
Submit a video that shows cubes are falling to the target after you run your code
Submit a .txt file that contain the log of the training, including the accumulated rewards for each episode, TD errors for each DQN update
Submit a .txt file with the following result:
How many times all the cubes fell successfully to the target in 100 trials if your state space consists only cup position?
How many times all the cubes fell successfully to the target in 100 trials if your state space consists both cup position and two cube’s position?