API Documentation¶
The safeopt package implements tools for Safe Bayesian optimization.
Stability verification¶
The Lyapunov class provides the main point of entry for the stability
analysis. It can be used to compute the region of attraction and together with
get_safe_sample() sets up the safe sampling scheme.
Lyapunov(discretization, lyapunov_function, …) |
A class for general Lyapunov functions. |
get_safe_sample(lyapunov[, perturbations, …]) |
Compute a safe state-action pair for sampling. |
smallest_boundary_value(fun, discretization) |
Determine the smallest value of a function on its boundary. |
get_lyapunov_region(lyapunov, …) |
Get the region within which a function is a Lyapunov function. |
Approximate Dynamics Programming¶
We use approximate dynamics programming to compute value functions.
PolicyIteration(policy, dynamics, …[, gamma]) |
A class for policy iteration. |
Functions¶
These are generic function classes for convenience. They are all compatible
with Lyapunov and PolicyIteration and can be added,
multiplied, and stacked as needed.
GridWorld(limits, num_points) |
Base class for function approximators on a regular grid. |
FunctionStack(functions[, name]) |
A combination of multiple 1d (uncertain) functions for each dim. |
Triangulation(discretization, vertex_values) |
Efficient Delaunay triangulation on regular grid. |
PiecewiseConstant(discretization[, …]) |
A piecewise constant function approximator. |
LinearSystem(matrices[, name]) |
A linear system. |
QuadraticFunction(matrix[, name]) |
A quadratic function. |
Saturation(fun, lower, upper[, name]) |
Saturate the output of a DeterministicFunction. |
NeuralNetwork(layers, nonlinearities[, …]) |
A simple neural network. |
GaussianProcess(gaussian_process[, beta, name]) |
A GaussianProcess model based on gpflow. |
GPRCached(\*args, \*\*kwargs) |
Create a new Mock object. |
sample_gp_function(discretization, gpfun[, …]) |
Sample a function from a gp with corresponding kernel within its bounds. |
Utilities¶
These are utilities to make working with tensorflow more pleasant.
utilities.combinations(arrays) |
Return a single array with combinations of parameters. |
utilities.linearly_spaced_combinations(…) |
Return 2-D array with all linearly spaced combinations with the bounds. |
utilities.lqr(a, b, q, r) |
Compute the continuous time LQR-controller. |
utilities.dlqr(a, b, q, r) |
Compute the discrete-time LQR controller. |
utilities.ellipse_bounds(P, level[, n]) |
Compute the bounds of a 2D ellipse. |
utilities.concatenate_inputs([start]) |
Concatenate the numpy array inputs to the functions. |
utilities.make_tf_fun(return_type[, …]) |
Convert a python function to a tensorflow function. |
utilities.with_scope(name) |
Set the tensorflow scope for the function. |
utilities.use_parent_scope(function) |
Use the parent scope for tensorflow. |
utilities.add_weight_constraint(…) |
Add weight constraints to an optimization step. |
utilities.batchify(arrays, batch_size) |
Yield the arrays in batches and in order. |
utilities.get_storage(dictionary[, index]) |
Get a unique storage point within a class method. |
utilities.set_storage(dictionary, name_value) |
Set the storage point within a class method. |
utilities.unique_rows(array) |
Return the unique rows of the array. |
utilities.gradient_clipping(optimizer, loss, …) |
Clip the gradients for the optimization problem. |