Page 337 - Practical Design Ships and Floating Structures

312

this case is called N-dimension space.
The selection of compartment layout design case can be described as following. Firstly use the case
chain in the indexing object class to find relative indexing object class. Nexl, according to the
objective of compartment layout design, finding out corresponding indexing object class. Then solving
the case record in this object class structure, in another word, to find all relative index. Because these
index object record all corresponding compartment layout design case, the relative case can be found.
In these relative cases, the most similar case can be found via evaluating the degree of similitude
between the new layout design task and the relative cases.
A compartment layout design task includes three parts. These are the objectives of layout design,
initial qualification given and restriction must be gratified. In the compartment layout design process,
the interaction of each sub-objective in the layout design task must be taken into account. In
compartment layout design process, many objectives and many restrictions must be meet at
simultaneity, and each objective and restriction belong to different property space.
Presently, there are three methods of case indexing based on CBR technology. They are Euclid
distance, Manhattan distance and infnite distance. All the three methods of case indexing concern the
weightiness of each property in the design case. In the process of solving problem of ship compartment
layout design, determining weightiness of compartment layout design property is even more difficult.
Based on foregoing analysis, this paper put forward the arithmetic to determine degree of similitude of
ship compartment layout design case based on neural network. It is an indexing model, which look on
a design task of compartment layout design as layout case index.
In the research of artificial neural network, BP(Back Propagation) arithmetic is widely studied and
utilized. For determining degree of similitude of compartment layout design case, a BP network which
have three layers and one output unit is constructed (see in Fig. 2).

output unit

Figure 2: A three-layer BP network

InFig.2, theinputunit I composedof Z,and I,, I= {Z, I, 1, and I,={ Ij , i= 1, 2, .-,
k}, I,= { Zj , i= I, 2, -*, k), W, is the weightiness matrix between output layer and hidden layer,
and W, is the weightiness matrix between input layer and hidden layer. BP is a recursive gradient
arithmetic to minimize the unbiased variance between the actual output and the predicted output in the
multi-layer BP network. The activation finction of processing units is S type finction, so the output of
units express in Eqn. (1):

In Eqn. (I), Si present the weightiness sum of all input in certain unit, Oi present the output of that
unit.
The aim to construct neural network is to acquired knowledge that ship design expert used in
determining degree of similitude of ship compartment layout design case. In actual it is using neural
network to present weightiness between sub-objective and properties in the design task of
compartment layout design case, which is the rate of contribution to degree of similitude of ship