Traditionally, the performance of multi-agent deep reinforcement learning algorithms are demonstrated and validated in gaming environments where we often have a fixed number of agents. In many industrial applications, the number of available agents can change at any given day and even when the number of agents is known ahead of time, it is common for an agent to break during the operation and become unavailable for a period of time. In this paper, we propose a new deep reinforcement learning algorithm for multi-agent collaborative tasks with a variable number of agents. We demonstrate the application of our algorithm using a fleet management simulator developed by Hitachi to generate realistic scenarios in a production site.

The problem of automated planning is known to be intractable in general. Moreover, it has been proven that in some cases finding an optimal solution is much harder than finding any solution. Existing techniques have to compromise between speed of the planning process and quality of solutions. For example, techniques based on greedy search often are able to obtain solutions quickly, but the quality of the solutions is usually low. Similarly, adding macro-operators to planning domains often enables planning speed-up, but solution sequences are typically longer. In this paper, we propose a method for optimizing plans with respect to their length, by post-planning analysis. The method is based on analyzing action dependencies and independencies by which we are able to identify redundant actions or non-optimal sub-plans. To evaluate the process we provide preliminary empirical evidence using benchmark domains.