| Definition | A customizable grid-world environment with relative positioning, reward clipping, and action space constraints. |
| No of environments | 2 |
The hurry_taxi environment builds on a grid-world framework, offering several customization options to enhance reinforcement learning tasks. One key feature is the Relative Position Wrapper, which computes the relative position between the agent and its target. This allows the agent to receive information about how far it is from its goal, enabling more strategic decision-making. The environment also includes the ClipReward Wrapper, which clips immediate rewards to a specified range, ensuring stable learning and avoiding extreme reward values that could destabilize training. Additionally, the DiscreteActions Wrapper restricts the action space to a finite subset, providing more control over agent behavior, particularly in more constrained scenarios.
The ReacherRewardWrapper allows for further customization by enabling the weighting of reward terms, which can be particularly useful when balancing multiple objectives in tasks that involve reaching specific locations or achieving certain goals. These features together make hurry_taxi highly versatile, allowing researchers and developers to experiment with different settings in grid-world environments. The project is open for contributions, with clear instructions on forking, setting up the project, and submitting updates or improvements to both the code and documentation.