Анонс открытой лекции по прикладной математике

ABSTRACT

Our talk is devoted to the control theoretical principles and algorithmic implementation of the advanced numerical methods from the Optimal Control Theory (OCT). We consider OCT techniques in the Reinforcement Learning (RL) computational framework. Recently some researchers (mostly from USA, China) have detected a deep interconnection between the classical OCT and the conceptual part of the RL.

We also refer to [2 - 3] for the related ideas and results. It is common knowledge, that the conventional OCT includes many effective, rigorous, and practically approved systems optimization algorithms (see e.g. [1] and [3]). This fact is due to a relatively large “uptime” of this specific control engineering area.

The main idea of the proposed concept consists in the natural methodology transfer from OCT to the modern RL. We discuss here the above transferring approach and indicate some novel RL research directions.   

Bibliography

[1] V. Azhmyakov, A Relaxation Based Approach to Optimal Control of Switched Systems, Elsevier, Oxford, 2019.

[2] D. P. Bertsekas, Value and Policy Iteration in Deterministic Optimal Control and Adaptive Dynamic Programming, Lab. for Information and Decision Systems Report LIDS-P-3174, MIT, May 2015 (revised Sept. 2015);

IEEE Transactions on Neural Networks and Learning Systems, Vol. 28, 2017, pp. 500-509.

[3] D.P. Bertsekas, Reinforcement Learning and Optimal Control, Athena Scientific, 2019