The master's thesis deals with the problem of multi-agent path finding for a group of transport vehicles, where the environment is presented with a road map and then converted into a state transition graph. The graph is searched using the graph search algorithms recently published in the literature, which are conflict resolution based, e.g., CCBS and CBS. In addition to executing the algorithm we also suggest an upgrade for a more efficient conflict detection and resolution.
The CCBS (Continuous Conflict-Based Search) algorithm is a two-level algorithm that searches for conflicts between agents, while taking their geometric shape into account. The CCBS algorithm works in a continuous time framework and therefore provides a more optimal solution than, e.g., CBS (Conflict-Based Search), which works in a discrete time framework. The CCBS uses the SIPP (Safe Interval Path Planning) algorithm on the low level, which is an upgrade of the A* algorithm, as it also considers the safe node and connection intervals while searching for the optimal path. On the high level it searches for the best node in the constraints tree. Nodes are created in the constraints tree with each new conflict discovered between agents.
The efficiency and speed of the CCBS algorithm also depend on conflict detection. For this purpose, we have presented two different types of conflict detection that work together. The first type uses adaptive sampling of the agents' path and detects all kinds of conflicts, including those that occur in different nodes and connections; however, the sampling makes it computationally demanding. The second type checks the occupancy intervals of the agents' nodes and connections and is computationally much less demanding, but only detects conflicts in the same nodes and connections. Due to being computationally demanding, the practical use of the CCBS algorithm is restricted to a smaller number of agents and to simpler or smaller state transition graphs, just as similar optimal algorithms. Namely, its computational complexity increases exponentially with increasing numbers of agents.
|