API Specification

This page covers the API specification for the following classes:

• lgrt4gps.lgrtn

• lgrt4gps.btn

Locally Growing Random Tree Node

class lgrt4gps.lgrtn.LGRTN(dx, dy, kerns=(), GP_engine='sklearn', div_method='center', wo_ratio=100, max_pts=100, inf_method='moe', optimize_hyps=False, lazy_training=False, sct=None, multi_processing=False, bounds=None, **kwargs)

Locally Growing Random Tree Node inherits from BTN

New in version 0.0.1.

Attributes

dxint

input dimension

dyint

output dimension

kernelslist (length: dy)

list of kernel instances (one for each output dimension), read only

GP_enginestr, optional

Default is ‘’ for simple high-speed GP or ‘GPy’ using the GPy package

div_methodstring, optional

Method to divide a dataset. Choices are ‘median’, ‘mean’, ‘center’

wo_ratiofloat, optional

width/overlap ratio. must be > 1

max_ptsint, optional

maximum number of points per leaf

inf_methodstring, optional

Method to combine local predictions Choices ‘moe’ (Mixture of Experts)

optimize_hypsbool

Turn hyperparameter optimization on or off

lazy_trainingbool

Wait with training until prediction is called

multi_processingbool

enables parrallel processing (for 1e5 datapoints overhead is too big)

Methods

add_data(x, y)	Adds data points to the node (or its children)
check_parent_child()	Checks if all parent relations are set
fit([optimize_hyps])	Prepares the model for prediction
get_child_xy()	Recursively gathers training data from all leaves
get_root()	Iteratively finds root of the tree
pprint([index, delimiter])	Pretty-print the binary tree.
predict(xt[, return_std])	Predicts posterior mean and standard deviation based on all leaf GPs
validate()	Check if the binary tree is malformed.

property X

input training data, read only

Returns

Xnumpy array (ntr x dx)

input training data

property Y

output training data, read only

Returns

Ynumpy array (ntr x dy)

output training data

_compute_musig(xt, return_std)

Calls gp from GPy to make predictions

Parameters

xtnumpy array (n x dx)

test inputs

return_stdbool (default: False)

whether std dev is computed or not

Returns

munumpy array (n x dy)

posterior mean

signumpy array (n x dy)

posterior variance

_distribute_data(x, y)

Distributes data to children

Parameters

xnumpy array (n x dx)

ynumpy array (n x dy)

_divide(x, y)

The current node grows two children and distributes its data

The node grows two children and distributes the new data (x,y) and its own data to the children

Parameters

xnumpy array (n x dx)

ynumpy array (n x dy)

_get_divider(x)

Computes parameters required for division of data set

1. Computes dimension with widest spread

2. Computes division values ‘median’, ‘mean’, or ‘center’

3. Computes overlap region Parameters ———- x : numpy array (n x dx)

Returns

dimint (1…dx)

dividing dimension

valfloat

dividing value

ofloat

overlap

_predict_local(xt, log_p, return_std)

Local prediction function

If current node is leaf, compute posterior estimates. If is not leaf, recurse further if any child has non-zero probability

Parameters

xtnumpy array (n x dx)

input test points

log_p :

log probability to reach the leaf

return_stdbool (default: False)

whether std dev is computed or not

_prob_left(x)

Computes probabilities of for left child

Parameters

xnumpy array (n x dx)

Returns

prob_left: numpy array (n,)

probabilities for left child

_setup_gps(optimize_hyps=None)

Prepare gp models (one for each output)

add_data(x, y)

Adds data points to the node (or its children)

Three options exist: 1. node is not a leaf: distribute data to children 2. node is a leaf, but full: grow two children and distribute 3. node is a leaf is not full: add data to node

Parameters

xnumpy array (n x dx)

input data to be added

ynumpy array (n x dx)

output data to be added

fit(optimize_hyps=None)

Prepares the model for prediction

For lazy training it enforces generation of GPs and trains the hyperparameters if needed

get_child_xy()

Recursively gathers training data from all leaves

Returns

childXnumpy array (ntr x dx)

Input training data concatenated from all children

childYnumpy array (ntr x dy)

Output training data concatenated from all children

property gps

list of GP instances (one for each output dimension), read only

Returns

gpslist

list of GP instances

property is_full

True if current node reached maximum capacity

Returns

is_fullbool

property kernels

list of kernel instances (one for each output dimension), read only

Returns

kernelslist

predict(xt, return_std=False)

Predicts posterior mean and standard deviation based on all leaf GPs

Starts the recursive call to prediction functions of leaves and combines the predictions according to the chosen inference method

Parameters

xtnumpy array (n x dx)

test inputs

return_stdbool (default: False)

whether std dev is computed or not

Returns

——-

munumpy array (n x dy)

posterior mean

signumpy array (n x dy)

posterior standard deviation

Binary Tree Node

class lgrt4gps.btn.BTN(parent=None, value=0, **kwargs)

Binary Tree Node inherits from Node in package binarytree

New in version 0.0.1.

Attributes

_parentBTN

parent

strstr

string to describe the node

Methods

check_parent_child()	Checks if all parent relations are set
get_root()	Iteratively finds root of the tree
pprint([index, delimiter])	Pretty-print the binary tree.
validate()	Check if the binary tree is malformed.

check_parent_child()

Checks if all parent relations are set

Returns

correctbool

True if relations of all subtrees are correct False if there is a missing link

get_root()

Iteratively finds root of the tree

Returns

rootBTN

root of the tree

property is_leaf

Returns True if current node is a leaf, False otherwise

Returns

is_leafbool

property is_root

Returns True if current node is a root, False otherwise

Returns

is_rootbool

property parent

Get parent of current node

Returns

parentBTN