Reinforcement learning

Learn to act by trial and error: model, plan, explore, learn

6 exercises

Each exercise is graded automatically, in your browser. The full statements are available once you have an account.

1. What this is for in real life

   Nobody tells a player which move to make: they learn by playing, and by seeing what pays off. Reinforcement learning gives machines this way of learning: no labels, only rewards, often delayed. - Games. In 2016…

2. The robot's logbook

   Your school's robotics club has a delivery robot that moves around a gridded warehouse. Before teaching it to move well, the team wants to give it a logbook: for every situation it can be in and every order it can…

3. The best route

   The robot's logbook is ready (previous exercise) and the club team has converted it into NumPy arrays. Now it has to be used: in each situation, which order should be given so that the robot gains as much as possible…

4. The spare-parts dispensers

   The robotics club's workshop gets its spare parts from second-hand vending machines. Each try costs a token; a dispenser gives a usable part with a certain probability, specific to it and unknown to anyone. The…

5. The test-run logs

   Other clubs lend their warehouses to the robot for demonstrations. For those warehouses, nobody gives you the map: you are only sent the log of a test robot that moved around in them, receiving random orders for a…

6. The cart and the pole

   For the school open day, the robotics club is preparing a demonstration. A small motorized cart rolls on a 4.8 m rail; a pole is attached to the cart by a free pivot. At rest, it hangs down. The task is to swing it up…