Size issue with rewards_df

[marota]

In

Oracle4Grid/oracle4grid/core/reward_computation/run_many.py

Line 27 in 1810f2d

data.append({"action": run.action, "timestep": t, "reward": run.rewards[t], "overload_reward": run.other_rewards[t]["overload_reward"]})

Contrary to run.rewards, run.other_rewards does not have the length of rewards_df.shape[0] if the configuration did not run until the end of the scenario.

Changing "for t in range(run.rewards.shape[0])" "for t in range(run.nb_timestep)" is probably not enough as a new issue raises later

[marota]

@mjothy Reproductible with
input_data_test_otherReward.zip

python main -d 1 -f oracle4grid/ressources/actions/wcci_test/agg_unitary_actions.json -e some_folder/input_data_test_otherReward -c 0 -as 249524660 -es 508848152
with timestep=50 and depth=1

config 'sub-1-6' stops running at timestep=45

[NMegel]

The problem comes from grid2op Runner (in the post-release version that we are using in Oracle for the moment)
The convention between res.rewards and res.other_rewards is not the same in case there is a game over within the simulation (one gets NaN, one gets its array shortened to the divergence timestep)

To handle this situation, a temporary fix has been set in run_many.py in commit 276fa69

Some harmonisation has been asked in Grid2op, with the following reproductible code

reproductibility_grid2op.zip

[NMegel]

An issue has been raised in grid2op repo
Grid2op/grid2op#170