Although users may do whatever they like to design and try their algorithms. We recommend wrapping a new algorithm as an Agent class as the example agents. To be compatible with the toolkit, the agent should have the following functions:
step: Given the current state, predict the next action.eval_step: Similar tostep, but for evaluation purpose. Reinforcement learning algorithms will usually add some noise for better exploration in training. In evaluation, no noise will be added to make predictions.