Alexey Skrynnik

In this work we study the behavior of groups of autonomous vehicles, which are the part of the Internet of Vehicles systems. One of the challenging modes of operation of such systems is the case when the observability of each vehicle is limited and the global/local communication is unstable, e.g. in the crowded parking lots. In such scenarios the vehicles have to rely on the local observations and exhibit cooperative behavior to ensure safe and efficient trips. This type of problems can be abstracted to the so-called multi-agent pathfinding when a group of agents, confined to a graph, have to find collision-free paths to their goals (ideally, minimizing an objective function e.g. travel time). Widely used algorithms for solving this problem rely on the assumption that a central controller exists for which the full state of the environment (i.e. the agents current positions, their targets, configuration of the static obstacles etc.) is known and they cannot be straightforwardly be adapted to the partially-observable setups. To this end, we suggest a novel approach which is based on the decomposition of the problem into the two sub-tasks: reaching the goal and avoiding the collisions. To accomplish each of this task we utilize reinforcement learning methods such as Deep Monte Carlo Tree Search, Q-mixing networks, and policy gradients methods to design the policies that map the agents' observations to actions. Next, we introduce the policy-mixing mechanism to end up with a single hybrid policy that allows each agent to exhibit both types of behavior - the individual one (reaching the goal) and the cooperative one (avoiding the collisions with other agents). We conduct an extensive empirical evaluation that shows that the suggested hybrid-policy outperforms standalone stat-of-the-art reinforcement learning methods for this kind of problems by a notable margin.

We present Hierarchical Deep Q-Network (HDQfD) that took first place in the MineRL competition. HDQfD works on imperfect demonstrations and utilizes the hierarchical structure of expert trajectories. We introduce the procedure of extracting an effective sequence of meta-actions and subgoals from demonstration data. We present a structured task-dependent replay buffer and adaptive prioritizing technique that allow the HDQfD agent to gradually erase poor-quality expert data from the buffer. In this paper, we present the details of the HDQfD algorithm and give the experimental results in the Minecraft domain.