Data from: Multi-alternative decision making with non-stationary inputs

One of the most widely implemented models for multi-alternative decision-making is the multihypothesis sequential probability ratio test (MSPRT). It is asymptotically optimal, straightforward to implement, and has found application in modelling biological decision-making. However, the MSPRT is limited in application to discrete (‘trial-based’), non-time-varying scenarios. By contrast, real world situations will be continuous and entail stimulus non-stationarity. In these circumstances, decision-making mechanisms (like the MSPRT) which work by accumulating evidence, must be able to discard outdated evidence which becomes progressively irrelevant. To address this issue, we introduce a new decision mechanism by augmenting the MSPRT with a rectangular integration window and a transparent decision boundary. This allows selection and de-selection of options as their evidence changes dynamically. Performance was enhanced by adapting the window size to problem difficulty. Further, we present an alternative windowing method which exponentially decays evidence and does not significantly degrade performance, while greatly reducing the memory resources necessary. The methods presented have proven successful at allowing for the MSPRT algorithm to function in a non-stationary environment.

多备选决策领域应用最为广泛的模型之一，即为多假设序贯概率比检验（multihypothesis sequential probability ratio test, MSPRT）。该模型具备渐近最优性，实现流程简洁直观，且已被广泛应用于生物决策行为的建模研究中。然而，MSPRT的应用场景仅局限于离散（「基于试次的」）、非时变情境。与之相对，现实世界中的决策情境往往是连续的，且伴随刺激的非平稳性。在此类场景下，依靠证据累积运作的决策机制（如MSPRT），必须能够丢弃逐渐丧失相关性的过时证据。为解决这一问题，我们通过为MSPRT增设矩形积分窗口与显式决策边界，提出了一种全新的决策机制。该机制可在证据随时间动态变化时，实现对备选方案的选中与撤销选中操作。通过将窗口大小与问题难度相适配，模型性能得到了有效提升。此外，我们还提出了一种替代性的窗口化方法：该方法通过对证据进行指数衰减处理，在未显著降低模型性能的前提下，大幅缩减了所需的内存资源开销。经实验验证，本文提出的两类方法可使MSPRT算法在非平稳环境中正常运作，效果优异。