Multi-Agent Coordination. Amit Konar

Multi-Agent Coordination - Amit Konar


Скачать книгу
implication of NFLT is that the elevated performance of one EA, say A, over the other, say B, for one class of optimization problems is counterbalanced by their respective performances over another class. It is therefore practically difficult to devise a universal EA that would solve all the problems. This apparently paves the way for hybridization of EAs with other optimization strategies, machine learning techniques, and heuristics.

      Hence, apart from the RL, hybridization of the EAs is also an effective approach to serve the purpose of multi‐robot coordination in a complex environment. The primary objective of an EA in the context of multi‐robot coordination is concerned with the minimization of the time consumed by the robots (i.e. the length of the path to be traversed by the robots) for complete traversal of the planned trajectory. In other words, robots plan their local trajectory, so that robots shifted from given positions to the next positions (subgoals) in a time‐optimal sense avoiding collision with the obstacles or the boundary of the world map. The optimization algorithm is executed in each local planning step to move a small distance. Hence, cumulatively robots move to the desired goal position using the sequence of local planning. There are traces of literature on hybridization of the EAs.

      In the literature of MAQL, agents either converge to NE or CE. The equilibrium‐based MAQL algorithms are most popular for their inherent ability to determine optimal strategy (equilibrium) at a given joint state. Hu et al. identified the phenomenon of similar equilibria in different joint states and introduced the concept of equilibrium transfer to accelerate the state‐of‐the‐art equilibrium‐based MAQL (NQL and CQL). In equilibrium transfer, agents recycle the previously computed equilibria having very small transfer loss. Recently, Zhang et al. attempted to reduce the dimension of the Q‐tables in NQL. The reduction is done by allowing the agents to store the Q‐values in joint state–individual action space, instead of joint state–action space.

      The book includes six chapters. Chapter 1 provides an introduction to the multi‐robot coordination algorithms for complex real‐world problems, including transportation of a box/stick, formation control for defense applications and soccer playing by multiple robots utilizing the principles of RL, the theory of games, dynamic programming, and/or EA. Naturally, this chapter provides a thorough survey of the existing literature of RL with a brief overview of the evolutionary optimization to examine the role of the algorithms in the context of multi‐agent coordination. Chapter 1 includes multi‐robot coordination employing evolutionary optimization, and especially RL for cooperative, competitive, and their composition for application to static and dynamic games. The latter part of the chapter deals with an overview of the metrics used to compare the performance of the algorithms while coordinating. Fundamental metrics for performance analysis are defined to study the learning and planning algorithms.

      Chapter 2 offers learning‐based planning algorithms, by extending the traditional multi‐agent Q‐learning algorithms (NQL and CQL) for multi‐robot coordination and planning. This extension is achieved by employing two interesting properties. The first property deals with the exploration of the team‐goal (simultaneous success of all the robots) and the other property is related to the selection of joint action at a given joint state. The exploration of team‐goal is realized by allowing the agents, capable of reaching their goals, to wait at their individual goal states, until the remaining agents explore their individual goals synchronously or asynchronously. Selection of joint action, which is a crucial problem in traditional multi‐agent Q‐learning, is performed here by taking the intersection of individual preferred joint


Скачать книгу