A cognitive system is an intelligent and adaptive system. It can learn from its environment, be it virtual or in the physical world, and it uses artificial intelligence to choose the best actions to reach its objectives. The aim is to design and teach a cognitive system to adopt and achieve its owner’s objectives.
This is the website of Gavin B. Rens.
The website provides details of my research into Cognitive Systems, a field within Artificial Intelligence.
More specifically, my field of research can be summarized as Autonomous Decision-making under Uncertainty. My particular interests are
- knowledge representation,
- probabilistic belief change,
- planning under uncertainty and
- agent architectures.
My doctoral studies fall mostly under the category of knowledge representation. My supervisors and I developed the Stochastic Decision Logic (SDL) (Rens et al., 2015). SDL is a logic for specifying partially observable Markov decision process (POMDP) models and for reasoning about such models, even if the models are partial or incomplete. Entailment of arbitrary queries about sequences of actions and observations can be answered by employing the logic. A sound, complete and terminating decision procedure is provided.
After completing my doctorate, I started working on probabilistic belief change, both revision and update. That is, the research involved revision of beliefs due to receiving correct information at odds with an agent’s current beliefs (Rens et al., 2016, 2018), and update of beliefs due to a changing environment (Rens and Meyer, 2018). My work on revision involved the method called Lewis imaging, which deals with the zero priors problem that Bayesian conditioning cannot deal with. My work on update involves generalizing the state estimation function of POMDPs to deal with exogenous events.
I have not focused on planning under uncertainty to the extent of knowledge representation and probabilistic belief change, but it has formed part of my research since I started my academic career. For instance, as part of my Masters degree, I developed a POMDP planner by extending DTGolog (Rens et al., 2008; Rens, 2010). And in 2017 my postdoc supervisor and I published an article reporting on an agent architecture involving POMDP planning (Rens and Moodley, 2017).
The journal article mentioned above (Rens and Moodley, 2017) is the culmination of research into combining the belief-desire-intention (BDI) framework with the POMDP formalism. The Hybrid POMDP-BDI Agent Architecture (Rens and Moodley, 2017) recommends actions in real-time (online), builds up a library of policies generated (to reuse later), and manages multiple goals in a sophisticated manner. I have also published two workshop papers on knowledge management frameworks explicitly involving probabilistic belief change and implicitly assuming the presence of a planning module (Rens, 2016; Rens et al., 2017).
I shall continue developing my understanding of and contributing to the state-of-the-art of effective knowledge representation methods, probabilistic belief change techniques, and online planning under uncertainty. And, where possible, I shall combine findings in these areas into systems for autonomous decision-making and reasoning about the effects of actions and the epistemic consequences of sensor inputs.
Finally, I have a continued interest in Reinforcement Learning and would like to incorporate Reinforcement Learning techniques into learning of symbolic knowledge (e.g., as in Haeming and Peters (2013)).
Last week, I attended the 13th International Conference on Agents and Artificial Intelligence (ICAART). On 4 February I presented my work about Online Learning of Non-Markovian Reward Models. On 5 Feb. I chaired two technical sessions. On the last day, (Saturday, 6 Feb.) I could simply attend talks. There were several interesting papers (for me) …
From the middle of last year, I’ve been advising or co-advising five Masters students at KU Leuven. I’ll list the topics they are working on below. The goal is to define how an MDP can be modelled by a first-order ProbLog program and to make a program that can solve such an MDP by using …
This is the start of my third and last year at KU Leuven as a postdoc. Eventually, we published a paper about learning non-Markovian reward machines in an active learning setting. So, in an MDP, if the agent’s rewards have temporal dependencies, we learn a finite state machine that models this temporal reward behaviour. At …