Build your own agent

On this page we quickly describe the steps we take before training our agents

Checklist

• Decide on Observation-Space and create an ObservationManager, if the already implemented ones do not suffice
• Create Reward Function and implement a RewardManager, if nececssary
• Check if config is as you'd like, most important configs to check are configs/train.yaml and configs/rl_env/single_agent_env.yaml. Most importantly, set the managers in the single_agent_env.yaml:

defaults:
  - obs_manager: next_point_obs.yaml
  - reward_manager: next_point_rewards.yaml
  - termination_manager: timeout.yaml

Check your reward function

We highly suggest you check your reward function, using the testing-framework to avoid unncessary errors, or scaling issues, as errors in the reward functions are the most likely reason for lack of learning.

What to expect

When using a single environment and images in your observation-space, on a reasonably capable GPU, it tatkes around 5-10 hours to see results.