As the number of inputs to a fuzzy system increases, the number of rules increases exponentially. This large rule base reduces the computational efficiency of the fuzzy system. It also makes the operation of the fuzzy system harder to understand, and it makes the tuning of rule and membership function parameters more difficult. Because many applications have a limited amounts of training data, a large rule base reduces the generalizability of tuned fuzzy systems.

To overcome this issue, you can implement a fuzzy inference system (FIS) as a tree of
smaller interconnected FIS objects rather than as a single monolithic FIS object. These
*fuzzy trees* are also known as *hierarchical fuzzy
systems* because the fuzzy systems are arranged in hierarchical tree structures.
In a tree structure, the outputs of the low-level fuzzy systems are used as inputs to the
high-level fuzzy systems. A fuzzy tree is more computationally efficient and easier to
understand than a single FIS with the same number of inputs.

There are several fuzzy tree structures that you can use for your application. The following figure shows commonly used fuzzy tree structures: an incremental, aggregated, or cascaded structure.

In an incremental structure, input values are incorporated in multiple stages to
refine the output values in several levels. For example, the previous figure shows a
three-level incremental fuzzy tree having fuzzy inference systems $$FI{S}_{i}^{n}$$, where *i* indicates the index of a FIS
in the *n*th level. In an incremental fuzzy tree, *i* = 1, meaning that each level has only one fuzzy inference
system. In the previous figure, the *j*th input of the
*i*th FIS in the *n*th
level is shown as input $${x}_{ij}{}^{n}$$, whereas the *k*th output of the *i*th FIS in the *n*th level is
shown as input $${x}_{ik}{}^{n}$$. In the figure, *n* = 3, *j* = 1 or 2, and *k* = 1. If each
input has *m* membership functions (MFs), each FIS has a
complete set of *m*^{2} rules.
Hence, the total number of rules is *nm*^{2} = 3 ⨉
3^{2} = 27.

The following figure shows a monolithic (*n* = 1) FIS
with four inputs (*j*=1, 2, 3, 4) and three MFs (*m* = 3).

In the FIS of this figure, the total number of rules is *nm*^{4} = 1 ⨉
3^{4} = 81. Hence, the total number of rules in an incremental fuzzy tree is linear
with the number of input pairs.

Input selection at different levels in an incremental fuzzy tree uses input rankings based on their contributions to the final output values. The input values that contribute the most are generally used at the lowest level, while the least influential ones are used at the highest level. In other words, low-rank input values are dependent on high-rank input values.

In an incremental fuzzy tree, each input value usually contributes to the inference process to a certain extent, without being significantly correlated with the other inputs. For example, a fuzzy system forecasts the possibility of buying an automobile using four inputs: color, number of doors, horse power, and autopilot. The inputs are four distinct automobile features, which can independently influence a buyer’s decision. Hence, the inputs can be ranked using the existing data to construct a fuzzy tree, as shown in the following figure.

For an example that illustrates creating an incremental fuzzy tree in MATLAB^{®}, see the "Create Incremental FIS Tree" example on the `fistree`

reference page.

In an aggregated structure, input values are incorporated as groups at the lowest
level, where each input group is fed into a FIS. The outputs of the lower level fuzzy
systems are combined (aggregated) using the higher level fuzzy systems. For example, the
following shows a two-level aggregated fuzzy tree having fuzzy inference systems $$FI{S}_{{i}_{n}}^{n}$$, where *i _{n}* indicates the index
of a FIS in the

In this aggregated fuzzy tree, *i*_{1} = 1,2 and
*i*_{2} = 1. Hence, each level includes a different
number of FIS. The *j*th input of the
*i _{n}*th FIS is shown in the figure as input $${x}_{{i}_{n}j}$$, and the

In an aggregated fuzzy tree, input values are naturally grouped together for specific decision-making. For example, an autonomous robot navigation task combines obstacle avoidance and target reaching subtasks for collision-free navigation. To achieve the navigation task, the fuzzy tree can use four inputs: distance to the closest obstacle, angle of the closest obstacle, distance to the target, and angle of the target. Distances and angles are measured with respect to the current position and heading direction of the robot. In this case, at the lowest level, the inputs naturally group as shown in the following figure: obstacle distance and obstacle angle (group 1) and target distance and target angle (group 2). Two fuzzy systems separately process individual group inputs and then another fuzzy system combines their outputs to produce a collision-free heading for the robot.

For an example that illustrates creating an aggregated fuzzy tree in MATLAB, see the example Create Aggregated FIS Tree on the `fistree`

reference page.

In a variation of the aggregated structure known as *parallel
structure*
[1], the outputs of the
lowest-level fuzzy systems are directly summed to generate the final output value. The
following figure shows an example of a parallel fuzzy tree, where outputs of
`fis1`

and `fis2`

are summed to produce the final
output.

The `fistree`

object does not provide the summing node Σ. Therefore,
you must add a custom aggregation method to evaluate a parallel fuzzy tree. For an
example, see the "Create and Evaluate Parallel FIS Tree" example on the `fistree`

reference page.

A cascaded structure, also known as combined structure, combines both incremental and aggregated structures to construct a fuzzy tree. This structure is suitable for a system that includes both correlated and uncorrelated inputs. The tree groups the correlated inputs in an aggregated structure, and adds uncorrelated inputs in an incremental structure. The following figure shows an example of a cascaded tree structure, where the first four inputs are grouped pairwise in an aggregated structure and the fifth input is added in an incremental structure.

For example, consider the robot navigation task discussed in Aggregated Structure. Suppose that task includes another input, the previous heading of the robot, taken into account to prevent large changes in the robot heading. You can add this input using the incremental structure of the following diagram.

For an example that illustrates creating an aggregated fuzzy tree in MATLAB, see the "Create Cascaded FIS Tree" example on the `fistree`

reference page.

When you evaluate a `fistree`

object, it returns results for only the
open outputs, which are not connected to any FIS inputs in the fuzzy tree. You can
optionally access other outputs in the tree. For instance, in the following diagram of an
aggregated fuzzy tree, you might want to obtain the output of fis2 when you evaluate the
tree.

You can add such outputs to a `fistree`

object. You can also remove
outputs, provided that the fuzzy tree always has at least one output. For an example, see
the "Update FIS Tree Outputs" example on the `fistree`

reference page.

A `fistree`

object allows using the same value for multiple inputs. For
instance, in the following figure, `input2`

of `fis1`

and
`input1`

of `fis2`

use the same value during
evaluation.

For an example showing how to construct a FIS tree in this way, see the "Use Same Value
for Multiple Inputs of a FIS Tree" example on the `fistree`

reference page.

You can add or remove individual FIS elements from a `fistree`

object.
When you do so, the software automatically updates the `Connections`

,
`Inputs`

, and `Outputs`

properties of the
`fistree`

object. For an example, see the "Update Fuzzy Inference Systems
in a FIS Tree" example on the `fistree`

reference page.

Once you have configured the internal connections in your fuzzy tree, the next step is to tune the parameters of the tree. For an example, see Tune FIS Tree for Gas Mileage Prediction.

[1] Siddique, N., and H. Adeli.
* Computational Intelligence: Synergies of Fuzzy Logic,
Neural Networks and Evolutionary Computing*. Hoboken, NJ: Wiley,
2013.