Indirect reinforcement learning for incident-responsive ramp control

A centralised strategy named indirect reinforcement learning ramp controller (IRLRC) has been developed in this paper to deal with ramp control problems for the congested traffic caused by incidents. IRLRC is developed on the basis of Dyna-Q architecture, under which a modified asymmetric cell transmission model (ACTM) and the standard Q-learning algorithm are combined together. The simulation-based test shows that compared with the no controlled situation, IRLRC can reduce the total travel time up to 24%, which outperforms the direct reinforcement learning (DRL) method with a reduction of 18% after the same number of iterations. (C) 2013 The Authors. Published by Elsevier Ltd. Selection and/or peer-review under responsibility of Scientific Committee