Fuzzified deep reinforcement learning-based evacuation model with early warning capabilities of crowd disaster [EMBARGOED]

Abstract

Understanding human response to crowd emergency is extremely complex and it plays a significant role in engineering construction designs and crowd safety. Individual choices, reasoning, and behaviours can’t be fully described by equations or rule-based methods. Accordingly, this research proposes a neuro-symbolic approach for modelling agents with human-level capabilities of reasoning and performance in an emergency evacuation. The proposed neuro-symbolic approach combines deep reinforcement learning (DRL) with evaluative fuzzy logic function to address the challenges of large amounts of required data, time, and trials-and-errors for policy optimization and to handle the assumption of the reward functions that may not be practical in real scenarios. This neuro-symbolic model has the potential to deal with the complexity of the environment and decision-making process via DRL and enhances the cognitive and visual intelligence via an evaluative fuzzy function, which continuously evaluates agent actions during the training process to boost pedestrian active response to their surroundings, with full awareness of time, thereby, a human-level capacity of reasoning. Moreover, this proposed model optimizes the computational demands of DRL and enables faster learning of new situations. A crowd disaster can easily occur all around. Thereby, prediction of the critical conditions is significant to prevent terrible disasters. Earlier research on entropy-based prediction models was primarily based on either the density or velocity attributes. Additionally, they evaluated the crowd risk by determining the level of chaos or abnormal behaviour rather than predicting a crowd disaster. Therefore, this research also proposes a crowd Boltzmann entropy-based prediction model that integrates the local density with average local speed to identify the critical situations that may result in crowd disasters, empowering sufficient prediction of crowd critical conditions and accurate description of the nature of a crowd motion therefore enabling early preventative intervention. The findings indicate that the proposed neuro-symbolic model can produce behavioral patterns that align with real observations of crowd evacuation, such as laminar flow, stop-and-go flow, and crowd turbulence. On top of that, a new evacuation behavior is observed, as some pedestrians avoid congestion at the exit until the density reduces. Moreover, the proposed model illustrates a higher accuracy and much faster converge than the pure PPO model with substantially minimal training timesteps as little as 2 to 8 percentage. Results of the proposed entropy-based prediction model reveal that the critical conditions last the longest in the nearest area to exit with the highest average entropy 0.44, and the shortest in the farthest area to exit with the lowest average entropy of 0.29 which is consistent with crowd literature. Meanwhile, the reliability study records an increase of the mean and standard deviation of evacuation time from 39.7s, 1.06 to 155.09s, 7.39 as crowd size increases from 15 to 200 pedestrians which implies a rise of uncertainty. The proposed prediction model outperforms well-established prediction approaches such as crowd pressure and crowd flow. Therefore, this work can provide crowd authorities and construction engineers with insights into complex crowd behaviour and critical conditions to evaluate evacuation plans and make decisions and sustainable designs to ensure crowd safety.