env.observation_space.high does not define the true maximum observations in the dataset

[enjeeneer]

• bauwerk version: 0.21
• Python version: 3.8
• Operating System: macOS

Description

The hard-coded upper limit on env.observation_space is unrealistically high for the load and pv_gen variables (dims 1 and 2 of the array printed below) i.e.

obs_high = []
for key, _ in env.observation_space.items():
    obs_high.append(env.observation_space[key].high)

print(np.concatenate(obs_high))

[7.500e+00 3.403e+38 3.403e+38 1.000e+00 1.000e+00]

This becomes problematic when one attempts to normalise the observation space using the floor env.observation_space.low and ceiling env.observation_space.high.

Looking through the repo, it seems the max value might be the load_scaling_factor and solar_scaling_factor for the load and pv_gen variables respectively? I see that default_load_data.txt and default_solar_data.txt are in [0, 1], and the load and pv_gen observations are the values from the .txt file multiplied by the scaling factor. Thoughts?

[rdnfn]

Thanks a lot for raising this, very good points!

The very high observation space bounds is an issue that is separate from the scaling factors. These bounds are set like

gym.spaces.Box(
    low=0, high=np.finfo(np.float32).max, shape=(1,), dtype=np.float32
)

In the main SolarBatteryEnv class. The bounds are set to the highest possible value as I assume we don't know the actual bounds (as we potentially wouldn't in a real world scenario), but it may be worth changing - as obviously in Bauwerk itself we have access to this information.

The second issue is more concerning, I will add my findings in #19.

[rdnfn]

Good news: the suggested update to the obs spaces has been implemented and a normalisation wrapper added with v0.3.0. Thus closing this issue.