Aggregating Computers and Experts

人类专家的观点是令人羡慕的经验和明显的认知偏见的产物。经过校准的计算机模拟可以做出令人印象深刻的精确预测，尽管其准确性值得怀疑。鉴于专家和计算机都以不同的未知偏差和准确性回答问题，我们不清楚如何最好地汇总信息，以评估我们所知道的，以及我们将从计算机或专家的进一步输入中获得什么。现有的技术（如高斯过程）可以使用来自专家或计算机的数据，并在先前测量的数据点之间进行插值。这些技术依赖于输入参数波动可能导致输出数据波动的程度。这个概念可以用数学方法在“核”中捕捉到。可以使用最近开发的分布式数值贝叶斯技术（顺序蒙特卡罗采样器）来有效地搜索巨大的核空间：这些技术能够更好地利用分布式硬件，而不是现有的替代方案（例如，马尔可夫链蒙特卡罗）。当内核还捕获专家和计算机输出之间可能存在的偏差的程度时，这种有效搜索的能力至关重要。整合计算机输出和专家判断的问题与联合利华加速新产品开发的能力有关。这种综合将能够：确定请求额外专家投入和（或）进行额外计算机模拟的效用；估计存在的偏差；确定信任专家和依赖计算机模拟的表面保真度之间的最佳折衷方案，并最终确定科学家在其产品设计活动中接受和采用此类模拟的程度。联合利华将与该学生合作，识别和访问与该挑战的具体实例相关的数据，并了解如何将正在开发的技术部署到新产品的配方中。

Human experts have opinions that are the product of enviable experience and pronounced cognitive biases. Calibrated computer simulations can make impressively precise predictions albeit with questionable accuracy. Given that both experts and computers each answer queries with different unknown biases and accuracies, it is unclear how we make best aggregate the information to assess what we know and what we would gain from further inputs from either a computer or an expert. Techniques (e.g., Gaussian Processes) exist to use data from either an expert or a computer and to interpolate between previously measured data-points. These techniques rely on some notion of the extent to which fluctuations in input parameters can result in fluctuations in output data. This notion can be captured mathematically in the "kernel". It is possible to use recently-developed distributed numerical Bayesian techniques (Sequential Monte Carlo samplers) to efficiently search the huge space of kernels: these techniques are better able to exploit distributed hardware than pre-existing alternatives (e.g., Markov chain Monte Carlo). This ability to search efficiently is of paramount importance when the kernel also captures the extent to which biases can exist between the output of an expert and a computer. The problem of aggregating computer outputs and expert judgement is pertinent to Unilever's ability to accelerate the development of new products. Such aggregation would make it possible to: ascertain the utility of requesting additional expert input and/or running additional computer simulation; estimate the biases present; identify the optimal compromise between trusting the experts and relying on the apparent fidelity of the computer simulations and ultimately the level of acceptance and adoption of such simulations to the scientists in their product design activities. Unilever will work with the student to identify and access data pertinent to a specific instance of this challenge as well as to understand how the technology being developed could be deployed in the context of the formulation of new products.