Uncertainty, Action & Interaction: in Pursuit of Cognitive Information Processing

我的研究目标是为在现实生活环境中增强人类认知的深刻认知信息处理（CIP）系统的设计和实现做出贡献。CIP涉及感知、学习、推理和面对不确定性和噪音的强大互动能力。我计划在我之前的工作的基础上，详细阐述有效的算法，以解决从低到高水平的持续CIP系统的不同认知水平：(I)在静态世界中对不确定知识的推理；（ii）在不确定环境中随时间的推理；(3)在不确定的环境中决定做什么；（iv）在不确定的多智能体环境中决定做什么。在(i)的背景下，我们提出构建我们自己的先验，这样就可以将有向无环图结构覆盖为信念网络和潜在树。我们特别研究了隐藏单位受欢迎程度的后验分布的哪个表示是最好的。在（ii）的背景下，我们的工作将主要集中在贝叶斯视角下建模和过滤之间的相互作用。一个方向是应用灵活的贝叶斯非参数建模方法学习未知概率模型进行隐状态估计。另一个方向是设计递归贝叶斯滤波框架，以在线方式有效地更新贝叶斯非参数建模性能。对于级别（iii），我们将首先研究迁移学习在贝叶斯部分可观察马尔可夫决策过程（POMDP）背景下的情况，特别是(a)当多任务和生活学习用于学习POMDP时，以及(b)当单个任务的训练数据不足时。其次，我们研究了一种更具推测性的途径，即以动态决策网络（DDN）的形式构建贝叶斯POMDP，即由奖励完成的DBN，然后将其转化为推理问题。最后，对于（iv），我们将关注合作控制的几个方面，特别是代理之间的共识，如何在“潜在博弈”中形成，以及这些特定博弈如何与分布式资源分配以及本地和分布式控制律相关联。

My research vision aims to contribute to the design and implementation of profound cognitive information processing (CIP) systems for augmented human cognition in real-life environments. CIP involves the ability to perceive, learn, reason and interact robustly in face of uncertainty and noise. I plan to build on my previous work to elaborate efficient algorithms to tackle different cognitive levels from low to high-level sustaining CIP systems: (i) reasoning on uncertain knowledge in static worlds; (ii) reasoning over time in an uncertain environment; (iii) deciding what to do in uncertain environment; (iv) deciding what to do in an uncertain and multiagent environment. In the context of (i), we propose to construct our own priors so that one can cover directed acyclic graph structures as belief networks and latent trees. We particularly investigate which representation of the posterior distribution over hidden units' popularity is the best. In the context of (ii), our work will mainly focus on the interaction between modelling and filtering in a Bayesian perspective. One direction aims at applying flexible Bayesian nonparametric modelling methods to learn unknown probabilistic models for hidden state estimation. The other direction is to design a recursive Bayesian filtering framework to efficiently update Bayesian nonparametric modelling performance in an online manner. For level (iii), we will first investigate the case of transfer learning in the context of Bayesian partially observable Markov decision process (POMDP), particularly (a) when multi-task and life-learning are used for learning a POMDP and (b) when the training data from single task are not sufficient. Second, we investigate a more speculative avenue which consists in structuring Bayesian POMDP in the form of dynamic decision network (DDN), that is, a DBN completed by rewards, and then translate it to an inference problem. Finally and for (iv), we will focus on how several aspects of cooperative control, particularly consensus between agents, can be formulated in ``potential games'' and how these specific games can be pertained to distributed resource allocation as well as to local and distributed control laws.