More good news, with two manuscripts accepted for the upcoming World Congress on Computational Intelligence (WCCI 2020) to be held in Glasgow next summer. No official announcement as of yet, but the conference is likely to go virtual due to the current situation with Covid-19 (UPDATE 22/06: This will be a virtual conference.
- A Bayesian perspective on classical control, Baltieri M., Accepted at the World Congress on Computational Intelligence (WCCI 2020), Glasgow, UK, 2020
The first one is my first single-author submission and summarises a program of research for Bayesian methods in an area of control theory that has received less attention in the last few decades, classical control. On the one hand, connections between Bayesian inference and optimal control theory implementations are by now well established, and on the other, classical controllers are known to be special cases of optimal control under the right assumptions. In practice however, the implications of these connections have not been fully explored and their potential not entirely exploited. Here I sketch a few algorithmic advandages of Bayesian formulations of PID controllers to highlight how Bayesian frameworks, and in particular active inference, might lead to improvements that could answer, or at least better frame, various questions and issues in the field of control theory (cf. Åström, Karl Johan, and Tore Hägglund. “The future of PID control.” Control engineering practice 9.11 (2001): 1163-1175.)
- Scaling active inference, Tschantz A., Baltieri M., Seth A. and Buckley C., Accepted at the World Congress on Computational Intelligence (WCCI 2020), Glasgow, UK, 2020
The second one stems largely from the work of Alec Tschantz and highlights a concrete attempt to connect active inference to modern large scale techniques adopted in reinforcement learning. In particular, this implementation relies on the idea of amortized inference that drives most current implementations of inference methods in RL/ML. Amortized inference is often used in place of mean-field models in the context of variational inference. Such mean-field models normally assume a factorisation over each data point for the parametrised variational density used to approximate the real (and usually uncomputable) posterior describing a dataset. On the other hand, amortized inference utilises a single neural network for the approximation of the parameters of the variational distribution (cf. Variational Autoencoders). The structure of the network, and its (more) limited set of parameters imposes constraints on the shape of the variational distribution by limiting the overwhelming flexibility of having a set of parameters for each data point. This work provides a preliminary analysis of the connections (and differences) between established methods in RL and the emerging framework of active inference, showing immediate applications and improvements in standard RL benchmarks (e.g., pendulum and hopper tasks). For more of Alec’s work, check his most recent preprint: Tschantz, Alexander, et al. “Reinforcement Learning through Active Inference.” arXiv preprint arXiv:2002.12636 (2020).