ARUP; ISBN: 978-0-9562121-5-3 - CMBBE 2012 - Cardiff University

Fig. 2. Hand characteristic parameters (left and middle) and experiment example (right).
Table 3. Experiments carried out for each subject.
Experiment ID
E1 E2 E3 E4 E5 E6 E7 E8 E9 E10 E11 E12 E13 E14 E15 E16
Bottle B1 B1 B1 B1 B2 B2 B2 B2 B3 B3 B3 B3 B4 B4 B4 B4
Level L2 L2 L1 L1 L2 L2 L1 L1 L2 L2 L1 L1 L2 L2 L1 L1
Task T2 T1 T2 T1 T2 T1 T2 T1 T2 T1 T2 T1 T2 T1 T2 T1
3.2 Design of the Artificial Neural Network
An artificial neural network is composed of a large number of simple processing
elements, neurons, running in parallel. The system function depends on the network
architecture (number of layers and neurons), the number of synaptic connections and the
processing function for every neuron (activation function). The experimental knowledge
is stored as the strength of the neuron connections (synaptic weights) by means of a
learning process, similarly to what happens in the human brain. The method used to
iteratively compute these weights as the network receives new information (inputs and
desired outputs or targets) is called training algorithm. It is applied to train the network
to perform some particular task. The optimum weights are obtained by an optimization
technique, usually the calculation of the mean square error between the network outputs
and the targets, for data not used in the training phase. This is called the test phase.
There are two problems usually tackled by ANNs: pattern classification and functions
approximation. Multiple-layer networks are quite powerful for functions approximation,
which is the subject of this specific work. Also the connectivity plays an important role
in successfully solving the problem (see [18] for detailed information). Accounting for
the complex nature of our problem, we decided to use a multi-layer fully-connected
feed-forward network, able to learn complex relationships. The selected training
algorithm was the Backpropagation technique [18], which is the generalization of the
Least Mean Squares algorithm to multiple-layer networks. It minimizes the mean square
error and performs a supervised learning, where a set of examples of proper network
behavior are provided for learning.
We developed, trained and tested two different ANNs: 1) A 2-layer feed-forward
network with a hidden layer (100 neurons); 2) a 3-layer feed-forward network with 2
hidden layers (75 and 50 neurons). Over each ANN 3 different tests were performed:
Test B2: The training set was composed of data from all the subjects and tasks, but
from bottles B1, B3, and B4. The test set was composed of data from all subjects
and tasks but only from bottle B2, that were not used for training.
Test S4: The training set was composed of data from all the bottles and tasks, but