Aleksandar Pluškoski

Dr. Aleksandar Pluškoski


Research Focus & Metrics

Academic Career Length

1Year (2026-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
1 May 2026 in Electronics
Verified
Journal Article
Open access
PDF file
APAleksandar Pluškoski
Aleksandar Pluškoski
ICIgor Ciganović
Igor Ciganović
MJMilos Jovanovic
Milos Jovanovic
JVJelena Vasiljlevic
Jelena Vasiljlevic
Policy learning under delayed reward conditions remains a significant challenge for end-to-end reinforcement learning (RL) agents. The difficulty increases for problems that require long-term planning and the execution of multiple dependent subtasks. As a result, solutions based on a single monolithic policy often suffer from unstable training. One possible solution to this problem could be to delegate the long-term planning to a separate model. This paper presents an implementation comprising two models: a large language model (LLM) for long-term planning and an execution model that solves subtasks. The execution model was trained via distillation from multiple teacher models trained with RL on individual tasks. The results presented in this paper demonstrate the benefits of this approach. By delegating long-term planning to the LLM, the agent can solve more complex problems than end-to-end agents trained with the proximal policy optimization (PPO) algorithm.

19 July 2026 in Electronics
Verified
Journal Article
Open access
PDF file
ICIgor Ciganović
Igor Ciganović
APAleksandar Pluškoski
Aleksandar Pluškoski
MJMilos Jovanovic
Milos Jovanovic
JVJelena Vasiljlevic
Jelena Vasiljlevic
Reward sparsity is a considerable problem in reinforcement learning when solving complex problems. In an environment where the reward is given only after successfully solving a problem in multiple steps, the signals for optimization can be too weak, preventing the agent from learning. However, this can be solved by breaking the problem into an array of simpler tasks that the agent can learn to perform. This paper presents a possible solution that uses a mixture-of-experts architecture. Separate models are used for planning, routing expert outputs, and solving singular sub-tasks. The solution was implemented in the MiniGrid environment, which allows for the design of arbitrarily complex problems. The results presented in this paper show that a multi-model agent can solve more complex problems comprising multiple steps than a single reinforcement learning agent.