Here is the fishing gym. Work IN PROGRESS!

Environments

This repository provides OpenAI-gym class definitions for the fisheries management problem. See Creating your own Gym Environments for details.

Install the python fishing_gym module by cloning this repo and running:

python setup.py sdist bdist_wheel
pip install -e .

So far, we have:

Simple fishing model defined in a continuous state space of fish biomass, with:

• Discrete action space with three actions: maintain harvest level, increase harvest by 20%, decrease harvest by 20%
• Discrete action space with n > 3 actions: action is taken as quota, quota = action / n_actions * K
• Continuous action space, action = quota.

Examples

Examples for running this gym environment in several frameworks:

• keras-rl
  • DQN
• tensorflow/agents
  • DQN
• stable baselines
  • PPO

Note that different frameworks have seperate and often conflicting dependencies. See the requirements.txt file in each example, or better, the official doucmentation for each framework. Consider using virtual environments to avoid conflicts.

Theory

The optimal dynamic management solution for the stochastic fisheries model is a "constant escapement" policy, as proven by Reed 1979. For small noise, this corresponds to the same 'bang-bang' solution for the determinstic model, proven by Clark 1973. Ignoring discounting, the long-term harvest under the constant escapement solution corresponds to the Maximum Sustainable Yield, or MSY, which is the optimal 'constant mortality' solution (i.e. under the constraint of having to harvest a fixed fraction F of the stock each year), as demonstrated independently by Schaefer 1954 and Gordon 1954.

The biomass at MSY can trivially be solved for by maximizing the growth function $X_{t+1} = f(X_t)$. Discretizing the state space, the dynamic optimal harvest can easily be found by stochastic dynamic programming.

Here, we seek to compare the performance of modern RL methods, which make no a-priori assumptions about the stock recruitment function, to this known optimal solution (given the underlying population dynamics).