Reinforcement learning environments that use Gaussian Processes to provide the agent with an uncertainty measure.
In the scenarios where an RL agent does not know the state with full certainty, it still must be able to make informed decisions. We have designed a series of RL environments in which a model of the environment is made as the agent plays the game. This allows for the agent to have a measure of uncertainty between discrete observations which provide the ground truth.
In the GP_Water environments, the agent must heat/cool the a fixed amount of water to a certain point. To do so, it must observe the Temperature and Mass Fraction of the solid/liquid/gas phases and either add or remove energy.
| Environment | Description |
|---|---|
| GP_Water-v0 | Basic Version. State contains [T, Terr, M, Merr] to achieve a fixed target. |
| GP_Water-v1 | Target values for T and M are added to the state. Target can be moved. |
| GP_Water-v2 | Observations actions are split for T and M. Each observation now comes with a cost. |
| GP_Water-v3 | Energy added to action, can remove either T or M from state and reach a target. |
Requirements for the environments: numpy, GPy, matplotlib, gym
Addtional Requirements for interactive demo: Tkinter
Rory Coles
CLEAN group
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent