Sprint #8 (June 22)

[pagand]

Feedback:

• Reinforcement learning dataset (generate a unified dataset for different trips, used for offline application. The dataset is a list of dictionaries with observations, next observation, actions, reward, and termination keys. Maybe we can make that publicly available for reproducibility, since all the data are standardized)

• Correct the values of rewards2 between -1 to 1)

• Include fuel consumption as one of the features

• Use decoded values for one-hot encoded columns

• Model: Consider LSTM and / or GRU

• hyper parameter tuning for sequence length and prediction horizon, as the data recording frequency is 1 min

• The goal is to achieve high accuracy as the model will be used to generate synthetic data

• Incorporate more columns for output of intrinsic temporal features

Action items:

• Create documentation for the reinforcement learning dataset, explaining the removal of specific columns and the reasons behind their removal and correct the current version
• ASAP
• Build a working version of the fuel consumption prediction model with high accuracy
• Begin writing draft for the paper
• Consider reviewing transformers in huggingface https://huggingface.co/docs/transformers/index

[yim-fan]

• [x] Documentation for rl dataset: View at https://docs.google.com/document/d/1A_7xobpUvPaWUna2VUPDBYc2873TITyhUimmYX2Z4x4/edit week 7 report
• [x] working version of fc prediction model
• [x] writing draft of the paper
• [x] transformers and / or attention mechanism