Figure 3

Prior influence on reconstruction performance of GA. Results for GA reconstruction for one sampled network (N = 15), population size p = 500, number of iterations 1000, crossover/selection rate q = 0.3, mutation rate m = 0.8, γ = 1. As in figure 2, the sampled network was used as 'perfect prior knowledge'. Left: AUC values without prior (column BIC) and for various settings of λ. When λ was decreased, the AUCs increased. However, unlike the inhibMCMC example, AUCs did not approach a value of 1, giving evidence that the GA converges to a local optimum. AUCs for BIC score optimisation were bad, emphasising the need for prior knowledge inclusion for larger networks. Right: Likelihood and prior differences. Since most of the observed prior differences were zero, only the non-zero values are shown. For each setting of λ, the left box corresponds to the observed distribution of likelihood differences, the right box to the prior differences. In the BIC column, only the likelihood difference distribution is shown, since no prior was used in this case.