Reinforcement Learning with Neuroevolution

Deadline - May 25, 2025

Choose a few (two to three) environments from the gymnasium library (we used this library in the second labs) and solve them using neuroevolution. Specifically, your task is to integrate reinforcement learning and neuroevolution and test the result on the selected environments.

I recommend selecting some environments from the Classic Control or Toy Text groups, but if you’re looking for a challenge, you can also try some environments from the Box2D or MuJoCo groups (for example, environments like LunarLander, Hopper, and Swimmer should be solvable, but I definitely don’t recommend environments like Humanoid or Humanoid Standup).

As for the choice of algorithm, you can try NEAT (or any similar algorithm you find elsewhere), or you can simply evolve the weight vector of a fixed network topology; the choice is yours.

To evaluate the proposed algorithm, it needs to be run multiple times with different random seeds, then compute the average performance of the best individual in each generation over these multiple runs.

Submit your solution via email, which should include:

• List of environments you tried
• Description of the algorithm along with its settings (hyperparameters, etc.)
• Algorithm code + instructions for running it
• Graphs showing the performance of the best individual over generations