This paper focuses on the performance of the second-order polynomial-based response surfaces on the reliability of spatially variable soil slope. A single response surface constructed to approximate the slope system failure performance function G（X） （called single RS） and multiple response surfaces constructed on finite number of slip surfaces （called multiple RS） are developed, respectively. Single RS and multiple RS are applied to evaluate the system failure probability pf for a cohesive soil slope together with Monte Carlo simulation （MCS）. It is found thatpe calculated by single RS deviates significantly from that obtained by searching a large number of potential slip surfaces, and this deviation becomes insignificant with the decrease of the number of random variables or the increase of the scale of fluctuation. In other words, single RS cannot approximate G（X） with reasonable accuracy. The value of pc from multiple response surfaces fits well with that obtained by searching a large number of potential slip surfaces. That is, multiple RS can estimate G（X） with reasonable accuracy.
The determination of optimal values for three parameters required in the original particle swarm optimization algorithm is very difficult. It is proposed that two new parameters simulating the harmony search strategy can be adopted instead of the three parameters which are required in the original particle swarm optimization algorithm to update the positions of all the particles. The improved particle swarm optimization is used in the location of the critical slip surface of soil slope, and it is found that the improved particle swarm optimization algorithm is insensitive to the two parameters while the original particle swarm optimization algorithm can be sensitive to its three parameters.
