Multi-Agent Systems (MASs) have been utilized in various settings and frameworks, and have thus been successfully applied in many applications to achieve different goals. It has been shown that MASs are more cost-effective as compared to building a single agent with all the capabilities a mission may require. Moreover, the cost is not the only driving factor for the adoption of MASs, e.g., safety is another important aspect: Deploying a group of agents, in a harsh or extreme environment, instead of a human team decreases the safety risks. Furthermore, MASs offer more flexibility and robustness when compared to a single-agent solution. The flexibility comes from dividing resources into separate groups, while robustness comes from the fact that a critical error in one agent does not necessarily endanger the success of a mission. Note that a mission may have many different constraints and aspects, however, the most trivial case has a single agent and a single task.
These kinds of missions can be planned by a human operator, overseeing a mission, without the need for an automated planner. On the other hand, more complex missions, that are utilizing a large number of heterogeneous agents and tasks, as well as constraints (precedence, synchronization, etc.) are not that trivial to plan for a human operator. These complex problems pose a great challenge to making a feasible plan, let alone the best possible one. Moreover, the increase in the power of available computing platforms in robotic systems has allowed the utilization of parallel task execution. More specifically, it allowed for possible parallelism in sensing, computation, motion, and manipulation tasks. This in turn had the benefit of allowing the creation of more complex robotic missions. However, it came at the cost of increased complexity for the optimization of the task allocation problem. To circumvent these issues, an automated planner is necessary. These types of problems are notoriously difficult to solve, and it may take too long for an optimal plan to be found. Therefore, a balance between optimality and computation time taken to produce a plan become very important.
This thesis deals with the formal definition of two particular Multi-Robot Task Allocation (MRTA) problem configurations used to represent multi-agent mission planning problems. More specifically, the contribution of this thesis can be grouped into three categories.
Firstly, this work proposes a model to represent different problem configurations, also referred to as missions, in a structured way. This model is called TAMER, and it also allows the addition of new dimensions in a more systematic way, expanding the number of problems that can be described compared to previously proposed MRTA taxonomies.
Secondly, this thesis defines and provides two different problem formulations, in a form of Mixed-Integer Linear Problem formulation, of the Extended Colored Travelling Salesman Problem (ECTSP). These models are implemented and verified in the CPLEX optimization tool on the selected problem instances. In addition, a sub-optimal approach to solving these complex problems is devised. Proposed solutions are based on the Genetic Algorithm (GA) approach, and they are compared to the solutions obtained by state-of-the-art (and state-of-practice) solvers, i.e., CPLEX. The advantage of using GA for planning over classical approaches is that it has better scalability that enables it to find solutions for large-scale problems. Although those solutions are, in the majority of cases, sub-optimal they are obtained much faster than with other exact methods. Another advantage is represented in a form of "anytime stop" option. In time-critical operations, it is important to have the option to stop the planning process and use the sub-optimal solution when it is required.
Lastly, this work addresses the one dimension of the MRTA problem that has not caught much of the research attention in the past. In particular, problem configurations including Multi-Task (MT) robots have been neglected. To overcome the aforementioned problem, first, the cases in which task parallelism may be achieved have been defined. In addition, the distinction between physical and virtual tasks and their mutual relationship in terms of parallel task execution has been introduced. Two models have been proposed and compared. The first one is expressed as ILP and implemented in the CPLEX optimization tool. The other one is defined as a Constraint Programming (CP) model and implemented in CP optimization tools. Both solvers have been evaluated on a series of problem instances.