The results of the experiment are shown in Fig 
and Fig .
The graphs show the evolution of the population of profiles as user interests
change over time. The profiles have been grouped together based on their specialization.
The initial population in each of the graphs is identical. Fig
shows the effect of varying the rate of change of user interest. Fig
shows the effect of varying the retention rate of the population.
Users would typically provide positive and negative feedback to the agents, while using the real system. For this experiment, the same effect is simulated by assuming an imaginary user with a certain bias, which represents the change in favor of the new interest. The bias is a parameter introduced solely for the purpose of this experiment which measures the rate of change of user interest. Bias is defined as the amount by which the fitness of certain profiles increases and that of the other profiles decreases every generation. Thus, if the bias is 0.01 in favor of politics, the fitness of profiles retrieving politics articles increases by 1%every generation and the fitness of the others decreases by 1%. The effect of a real user providing consistent feedback is simulated by setting the appropriate bias for the simulated user.
The retention rate ()
is set to 0.9. Offsprings of profiles are assumed to specialize to the same
category as their parents. For simplicity, the crossover genetic operator is
not used since the category of the offspring would be difficult to determine.
Only the mutation operator is used. Figure
shows the evolution of the population for different biases. The bias is held
constant for the entire duration plotted. Each graph plots the proportion of
profiles in the population which specialize in business and politics articles,
respectively, in successive generations. When the bias is sufficiently high,
the selective pressure in favor of politics is much stronger and the constitution
of the population changes relatively quickly. On the other hand, when the bias
is lower, the change in the population is much more gradual and takes longer.
In this case, the problem of profiles being half interested in two different topics does not arise as before. This is because feedback for each topic goes to the appropriate specialist. A Business profile would retrieve Business articles and only receive feedback for those. When the user interest changes, it will receive less positive feedback and will eventually be eliminated. However, the case where a Business profile will receive feedback for a Politics document will not arise. Thus, a profile will not have to specialize in two topics - it will either specialize in one, or be eliminated from the population.
The effect of varying the retention rate is shown in Fig .
The bias in this case is fixed at 0.015. The higher the retention rate, the
more stable the population is likely to stay across successive generations and
more gradual will be the change. When the retention rate is high, the change
is very gradual since the maximum change across two generations is constrained.
Lowering the retention rate sets the stage for potentially large changes in
the population. In the figure, as the retention rate is lowered, the change
in the constitution of the population is much more drastic and discontinuous
when it takes place.
The advantage of lowering the retention rate is that the population can react much faster. The disadvantage is that the changes are much more erratic and less controllable. The actual value should be set to achieve an optimal combination.