Detail project architecture:

• load balancing / scalability solutions (eg: AWS ASG)
• how can analysis done on the spot instance affect the backend host (eg: productionizing models / analysis)

More to be added in the following weeks.

Use the function created at #6 to generate a dataset for 1 month's worth of data. This will allow the start of parallel linear regression testing for prediction.

TODO:

• Find a proper requirements analysis template
• Write the document and add it in the _docs/ folder

Wider scope datasets will use 1-day scope dataset generation.

Add more:

• existing articles about serverless
• existing solutions and particular characteristics

Sample data

Split data into 2d arrays, each row representing one day's worth of data.

• Sample day period.
• Sample week period.
• Sample month period.
• Sample year period.

Individual prediction

Train any model having the date as an input.

• Regression over day scope. Input: (hr, min).
• Regression over week scope. Input: (week day, hr, min)
• Regression over month scope Input: (month day, hr, min).
• Regression over year scope (year day, hr, min).

Ensemble

• Neural network 4-(3/5)-1 -> Direct output.
• Genetic algorithm for weight determination -> Weighted sum.
• Genetic programming for combiner expression.

We should define a better scope for this (creating a FaaS framework from scratch is overkill for a semester at university). As long as we leverage cloud environments and distributed systems it is in "my field" and I can help (I have experience with Python, AWS, Docker, and Kubernetes).

Dataset expansion

• Versioning

• Start versioning the dataset generator (#6)

• 1 day scope

• Create dataset for 1 day's worth of data. (#6)
• Add arbitrary peaks to the data. (#6)
• Add arbitrary polling periods to the data. (#6)
• Self-contain 1-day dataset generation in a parametrized function (so it can be reused). (#10)
• Self-contain 1-day timetick generation into a function (#15)

• 1 month scope

• Create dataset for 1 month's worth of data. (#12)
• Plot 1 month data on a 4 row - 7 col grid (calendar style). (#12)

• 1 year scope

• Create dataset of 1 year. (#19)

Regression research

• Add possibility to transpose data dependence (day -> (time, load)) -> (time -> (day, load)). (#22)
• Add arbitrary polling intervals (including uneven). (#6)
• Self determined weighted sum over most likely time series periods (day, week, month, year).

Document how to install the kubernetes and kubeless:

• Document or present given documentation on how to install minikube
• Document how to install kubeless
• Document how to install Prometheus
• Document how to create and use a function
• Document how to expose the metrics (and also how to query them)

Add the last timetick as an argument, along with polling intervals.

Add the following diagrams:

• class diagram (if possible)
• usecase diagram
• other necessary diagrams (as decided by us)

TODO:

• Add Load Prediction sample Jupyter notebook presented at W3 seminar (10/16/2018) to the _notebooks/ folder.