In this study, a learning-based task allocation approach is proposed in order to increase the overall system performance. For this purpose, Q-learning algorithm is preferred. Theoretically, Q-learning algorithm is defined on single-agent frame. The difficulties of scaling up the multi-agent Q-learning to multi-robot systems are investigated. Two major approaches of multi-agent Q-learning in literature, distributed learning and centralized learning, are examined. To combine the advantages of these approaches, Strategy-Planned Distributed Learning approach is proposed. An important problem that appears in the application of Q-learning algorithm in multi-robot domain is to define discrete and finite state and action spaces. To represent the continuous state space, three methods, Fixed-Interval Discrete State Space (FIDSS), Continuous State Space with Distribution Function (DFCSS) and Variable-Interval Discrete State Space (VIDSS), are proposed. The continuous state space is discretized by using a fixed resolution value in FIDSS, whereas the discretization process is realized by a sequential clustering-based approach in an adaptive manner in VIDSS. DFCSS method represents the continuous state space by distribution functions in continuous way. The effectiveness of proposed approaches on system performance are demonstrated by applications. |