Skip to main content

MiR@W: Inference and Control for Complex Dynamical Systems

26th November 2012

Organisers: Sergio Morales and Chris Oates (Complexity)

Programme pdf

1200-1210 Welcome Notes

1210-1310 Milica Gašić (Cambridge)
Statistical Dialogue Optimisation
Spoken dialogue systems have recently acquired large attention due to appearance of speech driven personal assistants such as Siri or Google Now. Traditionally, such systems depend on a set of predefined rules for every state that can occur in a dialogue. The partially observable Markov decision process (POMDP) has been proposed as an alternative model for dialogue that allows inference of dialogue state and automatic optimization of the system behavior. In this talk methods will be presented that make POMDP-based dialogue optimization scalable and suitable for real world dialogue systems.

1310-1400 Buffet Lunch

1400-1430 Wilfrid Kendall (Warwick)
Coupling and control
A classic method in probability is that of coupling: constructing two copies of a random process which are interdependent, and designing the dependence so as to facilitate analysis of the process. I will give a short survey of recent work on coupling, with links to stochastic control methods.

1430-1530 Ben Calderhead (UCL)
Differential Geometric MCMC Methods with Application to Bayesian Model Selection of Dynamical Systems
I will discuss the latest advances in Markov chain Monte Carlo methodology that exploit the natural underlying Riemannian geometry of many statistical models. Such algorithms automatically adapt to the local correlation structure of the model, providing a highly efficient means of performing Bayesian inference over complex systems. I will provide examples of Bayesian inference using these methods on a variety of challenging dynamical systems described by nonlinear differential equations. Finally, I will discuss Bayesian model selection over such models based on estimation of Bayes Factors using thermodynamic integration.

1530-1600 Simon Spencer (Warwick)
Causal inference for biochemical networks
In observation experiments it is impossible to distinguish between association and causation. To uncover causal relationships, interventions must be included in the experimental design. In complex systems, such as biochemical networks, there is frequently a high degree of association between interacting parts of the system. The aim of causal network inference is to untangle the causal structure behind these associations. In this study we developed a statistical model that captures the effect of inhibitors (an intervention) in a protein signalling network. We then used this model to perform causal network inference on protein microarray data from breast cancer cell lines. We were able to demonstrate that a causal inference approach increases the accuracy of the inferred networks.

1600-1630 Coffee Break

1630-1730 James Anderson (Oxford)
Stability and Performance Analysis of Large-Scale Dynamical Systems
In our everyday lives we take for granted the fact that certain “networked-systems” will function in a reliable and predictable manner. Typical systems include the electricity grid, transport networks and the Internet to name just a few. Such systems are typically nonlinear, of high state dimension and contain uncertain parameters. For these reasons analysing the stability and performance properties of this type of system is challenging from both a computational and analytic point of view. In this talk I will describe how a decomposition based framework in conjunction with convex optimization and sum of squares programming can be employed to interrogate these large-scale systems.

1730-1800 Sam Brand (Warwick)
Controlling an Epidemic during Outbreak using Vaccination: Exact and Approximate Methods
Vaccination is an indirect method for reducing epidemic severity aimed at both protecting vaccinated individuals and reducing the subsequent infectious pressure of the epidemic. The problem is how to best to react using the available vaccine resource to optimally reduce the number of expected cases due to a novel pathogen spreading in a stochastic manner through a heterogeneous population. Results using the classical dynamic programming methodology will be presented as well as an investigation into using approximate dynamic programming (ADP) methods to solve optimal vaccination allocation problems computationally insoluble via classical methods. The ADP methodology reduces to the inference of the unknown value function of the optimal control problem via recursive Monte Carlo. The sensitivity of this inference to the parametric functional model will be discussed.