| Definition | A set of fishing environments for reinforcement learning, focused on sustainable fishery management and dynamic ecosystem modeling. |
| No of environments | 10 |
The fishing environments provide various models to simulate sustainable fishery management using reinforcement learning. These models range from simple scenarios with discrete and continuous action spaces to more complex settings that include a tipping point in the ecosystem’s growth model and uncertainty in parameters. In each environment, the agent controls fishing quotas while navigating the dynamics of fish biomass, governed by parameters like the growth rate, carrying capacity, and process noise. Users can configure these parameters to create different scenarios, allowing for a wide range of experimental conditions in resource management and ecological sustainability.
The environments are designed to test the agent’s ability to discover optimal fishing policies, comparing its strategies to known theoretical solutions like Maximum Sustainable Yield (MSY). Based on the dynamic management theory by Reed (1979) and earlier models by Schaefer and Gordon, the optimal solution often involves a “constant escapement” policy, where the agent learns to balance the harvest with the fish population’s growth. This setup allows for benchmarking modern reinforcement learning methods against classic dynamic programming approaches, offering insights into their effectiveness in environmental resource management.
| Example observations | |
| Example actions | |
| Example reward |