Reinforcement learning for autonomous agents exploring environments: An experimental framework and preliminary results

Reinforcement Learning (RL) is being growingly investigated as an approach to achieve autonomous agents, where the term autonomous has a stronger acceptation than the current most widespread one. On a more pragmatic level, recent developments and results in the RL area suggest that this approach might even be a promising alternative to current agent-based approaches to the modeling of complex systems. This work presents an investigation of the level of readiness of a state-of-the-art model to tackle issues of orientation and exploration of a randomly generated environment, as a toy problem to evaluate the adequacy of the RL approach to provide support to modelers in the area of complex systems simulation, and in particular pedestrian and crowd simulation. The paper presents the adopted approach, the achieved results, and discusses future developments on this line of work.