INTELLIGENT WIDGET RECONFIGURATION FOR MOBILE PHONES: MINIMAL INTELLIGENCE ALGORITHM(1)

Test results show that MI is unable to accurately predict user activities. Prediction is based on the previous day data and the percentage of correct predictions is the lowest among all three algorithms. For the weekly repeating usage patterns (Figure 8), there are not many repeated activity patterns for a given context. This causes the MI algorithm to have low accuracy that tapers off at around 30% after 3 weeks. It is still able to achieve 30% because at any time, four different contexts are used for the learning process and this enables the algorithm to at least distinguish and identify to a certain extent, the context used.


Fig8Intelligent Widget_decrypted
Figure 8: MI Prediction Accuracy for weekly repeating usage pattern

For the daily repeating usage pattern, the MI algorithm is able to achieve higher accuracy than the previous two usage patterns as there are more repeated activities (Figure 9).
Fig9Intelligent Widget_decrypted
Figure 9: MI Prediction Accuracy for daily repeating usage pattern

SINGLE LAYER PERCEPTRON WITH ERROR CORRECTION
Test results for SLP show that it has the best overall performance among the three learning algorithms. SLP analyzes data preceding the prediction and does its calculations based on different weighted inputs. It does not try to recognize patterns in the input data in order to predict user activity. Instead, weights are readjusted based on the next day’s data. This removes any problem caused by conflicting input data which can cause inaccuracies in the bias. Results in Figures 10 and 11 indicate that it is better equipped than MI and MLP to predict user activity for both weekly and daily repeating datasets.

For weekly repeating data, SLP gets about 60% prediction accuracy compared to an approximately similar average for MLP (Figure 12). There is no significant difference over MLP. This shows the weakness of the SLP algorithm as it neither learns with the data nor attempts to recognize any data patterns. For highly regular usage patterns, it achieved over 90% prediction accuracy. Usage pattern consistency as with the set of daily repeating data offers inherent data stability to enhance its prediction accuracy.
Fig10Intelligent Widget_decrypted
Figure 10: SLP Prediction Accuracy for weekly repeating usage pattern
Fig11Intelligent Widget_decrypted
Figure 11: SLP Prediction Accuracy for daily repeating usage pattern MULTI LAYER PERCEPTRON WITH BACK PROPAGATION

Test results show that MLP has the similar performance to SLP when the usage pattern is regular as with the daily repeating data set (Figure 13). For weekly usage patterns, however, its performance generally trails SLP although the average is similar (Figure 10 and Figure 12). The main reason for this result is because MLP needs to learn from existing data. When the data does not exhibit a significant level of repeating usage patterns, conflicting trends may arise and cause learning errors.
Fig12Intelligent Widget_decrypted
Figure 12: MLP Prediction Accuracy for weekly repeating usage pattern