Cooperative System Identification

My colleague Inês Lourenço has just published a pre-print of her latest work: Cooperative System Identification via Correctional Learning. It has been developed in collaboration with me, Cristian Rojas and Bo Wahlberg.

In this work, we consider a scenario where a teacher knows a correct model of a system and aims to assist a student who is trying to estimate it. However, the teacher cannot directly transfer its model to the student – for example, the teacher’s knowledge might be abstract (“what is your model of how to drive a car?”).

We propose correctional learning as an approach to this problem: in order to assist the student, the teacher can intercept the observations collected from the system and modify them so as to maximize the amount of information the student receives.

Have a look at the pre-print, and send Inês or me any comments you may have!

TSP Published

Our paper Inverse Filtering for Hidden Markov Models with Applications to Counter-Adversarial Autonomous Systems has now been published on IEEE Xplore – have a look and let us know if you have any comments!

TSP Accepted

I am glad to announce that our paper Inverse Filtering for Hidden Markov Models with Applications to Counter-Adversarial Autonomous Systems (an updated version is available in Chapter 7 of my thesis) has been accepted for publication in the IEEE Transactions on Signal Processing.

In this work, we formulate and propose solutions to the problem of inverse dynamical Bayesian inference (filtering) for hidden Markov models. In other words, given a sequence of posteriors, we present identifiability results together with a method to identify the filter’s parameters, comprising the transition matrix of the hidden process, the matrix of observation likelihoods as well as the sequence of measured observations y₁, …, y_N.

As an application of our results, we demonstrate the design of a counter-adversarial autonomous system: How can we estimate the accuracy of an adversary’s sensor, based on measurements of its actions? We believe that this paper significantly advances the toolset available to practitioners dealing with counter-adversarial systems.

The work has been developed in collaboration with Cristian Rojas, Vikram Krishnamurthy and Bo Wahlberg. We are very excited about this line of research – let us know if you have any comments!

CDC'20 Accepted

I’m happy to announce that our paper How to Protect Your Privacy? A Framework for Counter-Adversarial Decision Making has been accepted for presentation at the 59th IEEE Conference on Decision and Control (CDC’20).

My colleague Inês Lourenço is first author of the paper, which has been developed in collaboration with me, Cristian Rojas and Bo Wahlberg. It is an extension of our recent works that investigate counter-adversarial signal processing.

Virtual ICML'20

The International Conference on Machine Learning (ICML’20) is almost over. Due to the current extraordinary global situation, the conference was fully digital this year.

We presented our paper Fast and Consistent Learning of Hidden Markov Models by Incorporating Non-Consecutive Correlations during two virtual poster sessions. Thanks everyone that “stopped by” our poster – the discussions were very constructive!

Successful Defense of PhD Thesis

Today, I successfully defended my PhD thesis Hidden Markov Models: Identification, Inverse Filtering and Applications against a committee consisting of

• Prof. John Lygeros (opponent)
• Prof. Jan H. van Schuppen
• Prof. Tobias Rydén
• Assist. Prof. Ayça Özçelikkale

and was awarded my PhD degree. 😃

ICML Accepted

I’m very happy to announce that our paper “Fast and Consistent Learning of Hidden Markov Models by Incorporating Non-Consecutive Correlations” has been accepted for this year’s International Conference on Machine Learning (ICML’20).

This work has been developed in collaboration with Cristian Rojas, Eric Moulines, Vikram Krishnamurthy and Bo Wahlberg.

In the paper, we study how the parameters of an HMM can be estimated from observed data in fast and algorithmically attractive ways. The work could benefit practitioners working with large-scale time-series datasets (e.g., bioscience, finance, social networks, …).

A “preprint” of the paper is available in my PhD thesis (see Chapter 4) – to be defended tomorrow.

PhD Thesis

My PhD thesis Hidden Markov Models: Identification, Inverse Filtering and Applications has been printed. If you would like a physical copy, let me know. You can download an electronic version here.

It will be publicly defended on the 2nd of June, with a committee consisting of

• Prof. John Lygeros (opponent)
• Prof. Jan H. van Schuppen
• Prof. Tobias Rydén
• Assist. Prof. Ayça Özçelikkale

You can join via Zoom or in person (Kollegiesalen) at 09:00.

How to Protect Your Privacy?

My colleague Inês Lourenço has just published a pre-print of her latest work: How to Protect Your Privacy? A Framework for Counter-Adversarial Decision Making.

It has been developed in collaboration with me, Cristian Rojas and Bo Wahlberg. It is an extension of our recent works that investigate counter-adversarial signal processing.

The central question asked in the paper is: How can a decision maker select an action that

1. ensures that its private belief is not exposed,
2. while minimizing its cost?

Have a look, and send Inês or me any comments you may have!

New Pre-Print (Inverse Filtering)

You can now find a pre-print of our latest work Inverse Filtering for Hidden Markov Models with Applications to Counter-Adversarial Autonomous Systems, in which we provide important extensions to the inverse filtering algorithms proposed in our earlier NeurIPS paper.

By observing, or intercepting, posterior distributions from a Bayesian filter, we seek to estimate i) the model of the dynamic system, ii) the accuracy of the sensors and iii) the measured observations.

We also discuss the design of counter-adversarial systems. As in our ICASSP’20 paper, an enemy:

1. measures our current state via a noisy sensor,
2. computes a posterior estimate (belief) and
3. takes an action that we can observe.

Based on observations of the enemy’s actions and knowledge of our own state sequence, we estimate the accuracy of the enemy’s sensors.

Have a look, and feel free to send me any comments you may have by email!