The AI Centre studies advanced approaches to the creation of autonomous intelligent agents and multi-agent systems capable of solving complex practical problems. Research is focused on the development of agents endowed with the ability to independently formulate and achieve goals, accumulate knowledge, and effectively utilise external tools and devices. Technologies for multi-agent systems are also actively being developed, allowing various agents to interact with one another, competing, collaborating, and evolving through improvement.
Analysis of the complexity of deep learning models and the investigation of algorithms in linear algebra, statistics, and optimisation, which play a key role in their development
Tasks
- Development of new methods for constructing confidence intervals in the task of autonomous reinforcement learning, particularly for methods using temporal differences.
- Application of the developed methods for offline evaluation of algorithms in the domain of recommendation systems.
- Utilisation of entropy-based reinforcement learning methods to enhance the diversity of generated reasoning chains in mathematical domain tasks.
Modelling time-independent and dependent processes and constructing their surrogates using artificial intelligence methods
Tasks
Development of reinforcement learning approaches with vector reward functions for non-trivial rewards, exemplified by the control of complex systems (using charged particle beams as an example)
Scalable contextual bandits and RL algorithms for personalised recommendation tasks
Tasks
Development of recommendation management algorithms in reinforcement learning tasks, as well as contextual bandits: methods of scalable contextual bandits based on measure substitution for large action spaces and methods for offline evaluation of reinforcement learning algorithms using the Fitted Q-evaluation algorithm in the context of discounted and averaged Markov Decision Processes (MDP)
Effects
Providing personalised and more relevant responses from artificial intelligence models
Development of an innovative software solution with a multi-agent architecture based on an adapted LLM
Tasks
Development of targeted agents for the fields of science, technology, and innovation (for example, for information retrieval of relevant data in subject areas, analysis, meta-analysis, or data synthesis), communication strategies between agents, methods for error handling and conflict resolution during agent interactions, a prototype software solution based on the adapted LLM, developed targeted agents, and open software modules
Effects
- Timely identification of promising niches for focusing the efforts of research teams
- Increase in breakthrough results and provision of proactive development of technological innovations