Tools
Tools
Djemai, Ramzi (2026) BiLSTM-guided neuro-symbolic semantic planning, re-planning, and backtracking for efficient emergency evacuation. Doctoral thesis, London Metropolitan University.
Indoor emergency evacuation under uncertainty requires planners that reason about building semantics, respond to dynamic hazards, and recover from routes that become infeasible during execution. Classical heuristic and dynamic search methods operate on purely metric representations and react to observed changes, but they do not exploit the semantic structure of the building and cannot anticipate when a chosen route is likely to fail. This thesis develops an ontology-guided neuro-symbolic framework that integrates a formal knowledge base with an incremental symbolic planner and a learned predictive component while preserving the guarantees of admissible heuristic search.
The symbolic layer is built on a knowledge base organised as topology, situations, actions, and events, expressed in OWL and refined through SWRL rules that derive admissible transitions and a semantic heuristic. A semantic variant of Lifelong Planning A* operates on the resulting state-transition graph and supports event-driven re-planning together with reactive backtracking when a dead-end is reached. A bidirectional LSTM with two heads is then introduced as a strictly subordinate learned layer. One head estimates residual cost and is combined with the semantic heuristic through a bounded blend that retains admissibility. The other head produces a calibrated backtracking probability that acts as a predictive gate on action selection without altering the set of admissible transitions.
The framework is evaluated on three structurally distinct building ontologies, a vertical tower, a multi-block hospital, and a multi-wing school, across a three-tier scenario taxonomy that varies hazard count and disruption severity. Seven planner configurations, including three blending weights of the proposed system and four established baselines, are compared on matched scenarios. The proposed system reduces mean evacuation time and node expansions relative to the strongest incremental baseline while maintaining success rate, and the observed improvements are consistent across the three buildings. The contribution is an integration pattern in which learned guidance accelerates symbolic search under uncertainty without eroding its formal properties, offering a reproducible basis for ontology-guided evacuation planning in dynamic indoor environments
 |
View Item |