Registration Number
0223U005290,
(0121U100534)
, R & D reports
Title
Methods and tools of reinforcement learning for the tasks of analysis and forecasting for complex systems
popup.stage_title
Методи та засоби навчання з підкріпленням для задач аналізу та прогнозування складних систем
Head
Kasianov Pavlo O., д.ф.-м.н.
Registration Date
19-12-2023
Organization
Educational and Scientific Complex "Institute for Applied System Analysis" of National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute"
popup.description1
Development of computationally efficient algorithms for reinforcement learning problems for Markov decision-making processes with constraints and infinite states and partially observed Markov processes. Development of new results for structural properties of approximate solutions of one-criteria solutions problems with a discounted criterion of expected total costs and average costs per unit time. Development of methods for calculation of the optimal and almost optimal strategy for single-criteria Markov decision-making processes and partially observed Markov decision-making processes with constraints.
Solve classes of reinforced learning tasks, including developing and training a decision-making system that is highly
indicators of success finds shorter ways to the set target points in scenes both from training, and from validation
samples, demonstrating the ability to generalize.
popup.description2
The investigation is devoted to the study of problems of controlled partially observable stochastic systems with discrete time, optimization of expected total discounted costs. New sufficient conditions regarding transition probability and observation for weak continuity of transitions for MDP with trust states are obtained, several well-known such conditions are generalized, and existence of an optimal strategy, validity of optimality equations determining optimal strategies, convergence of iterations to optimal values are established for POMDP. In addition, a software product was created in which, using deep reinforcement learning and long-term planning, a neural network was obtained. This network approximates a strategy that achieves better results and metrics on a selected data set.
Product Description
popup.authors
Babych Halyna
Kondratova Liudmyla
Kupenko Olha P.
Levenchuk Liudmyla
Marchenko Halyna
Paliichuk Liliia
Shubenkova Iryna A.
popup.nrat_date
2023-12-19
