Motivation¶

The Wall¶

There’s a wall between the optimization algorithm and problem when:

They are written in different language
- i.e. algorithm in python and problem in matlab
They are located at different place and can’t be easily put together
- i.e. remote optimization due to COVID-19
The policy doesn’t allow the environment configurations that are needed for the algorithm/problem
- i.e. The strict policy in the accelerator control rooms

Even within the algorithm there could be a wall: imagine that you want to use some python modules in your matlab algorithm.

Break the Wall¶

Calling one language script/function within another language script/function is cumbersome and sometimes not feasible. Translating the code to the target language usually requires expert skills and lots of attention, energy and time. Besides, the translation process is highly error-prone. Use an implementation that already exists is risky from time to time. The way to go is breaking the wall.

Introducing Teeport.

Concepts¶

Before officially introduce Teeport, let’s have a look at the basic concepts/terms used frequently in a Teeport context: Evaluator, Optimizer, Processor and Monitor.

Evaluator¶

An implementation of a optimization problem. Usually can be abstracted as a function:

Y = evaluate(X, configs)

Where the configs variable is the configuration of the problem, say:

configs = {
    'noise_level': 0.1,  # emulate the experimental env
    'wall_time': 1,  # emulate the experimental env
    ...
}

Optimizer¶

An implementation of an optimization algorithm:

x_best = optimize(evaluate, configs)

The evaluate variable is the target function to be optimized:

Y = evaluate(X)

An example of configs:

configs = {
    'D': 8, # dimension of the problem to be optimized
    'N': 300, # maximum iterations
    ...
}

Processor¶

A function that processes data and return the processed result:

result = process(data, configs)

An example of configs:

configs = {
    'return_grad': False,
    'ard': True,
    ...
}

Monitor¶

A visualization of the data flow between the Optimizer and the Evaluator that catches some feature of the data.

Philosophy of Teeport¶

Teeport is designed to break the wall between the optimization algorithms and problems.