- Collect all the rewards (multi-color bouncing squares)
- Dodge the enemies (red bouncing squares)
I created a game from scratch with the purpose of teaching a reinforcement learning agent to play it. The core of the game is the environment class that sets what is possible within its boundries and allows the objects within it to act in a certain way. The three main objects are Player, Enemies and Rewards, with rewards and enemies inheriting from the same base. In order to train an rl-agent, the visualization needed to be completely decoupled from the underlying simulation (unless you want to spend resources visualizing training).
This game can be played by a player or a reinforcement learning agent. You can either play the game by selecting the play mode or what the ai play with the ai mode. Just run python main.py --mode play --config game. The other config option is 'test' where all the rewards and enemies will remain static, which is helpful if you want to print your state to console and watch it as you move around.
Double tap "spacebar" to play the exact some environment that you just died on (or won), and double tap "q" to randomly create a new environment within the bounds you specified in the config.json.
The RL agent has been able to consistently beat the game. In the near future, I will write an article about the methodology and "tricks" I used to get the agent to learn the game. This agent is far from perfect and will be updated to more advanced algorithms such as actor-critic.
A lot of the work getting the RL agent to work is found in the StateTranslator class. This takes the positions returned by the environment at each step, and transforms them into rich features for the agent to learn at a much faster rate.