- Myers, May, Ricker, Beverton-Holt, Allen, ModelUncertainty, NonStationary - refactor code and tests to conform to standards, see Makefile [42]
0.0.4
Release notes
- fix plotting method - Add three new environments:
- `fishing_v2` uses an environment that has a tipping point (as case which technically violates the concavity assumptions of Reed's 1979 proof, though the optimal solution with small noise is still essentially the same constant escapement strategy) - `fishing_v3` with observation error (also not covered by Reed's proof, and the optimal strategy is not constant escapement in this case, as we showed recently using POMDP methods: https://github.com/boettiger-lab/pomdp-intro ) - `fishing_v4` adds uncertainty around the parameters. Technically I believe you can nest this case into a classic MDP framework using a Bayesian learning rule, see https://github.com/boettiger-lab/mdplearning . Curious to explore 'transfer learning' in this one as well, e.g. can it learn something more general / more robust than agents (like all the other examples) which are only ever trained on a fixed parameter set?
0.0.3
- adds simulate and plotting methods - some tweaks to the default parameter choices - updated stable-baselines3 examples
0.0.2
Initial release of `gym_fishing`: An OpenAI Gym implementation of the classic optimal fishing harvest problem, for benchmarking Deep Reinforcement Learning applications to ecological management and conservation problems.