1. Pre-Event Heuristics and Simulation A multi-agent system designed for event planning, layout optimization, and safety validation is presented below. The interface leverages multiple agents working in parallel to handle different responsibilities. Together, they ensure events are configured for efficiency, safety, and compliance. Agent Roles and Details: Information Agent: Collects and structures data (venue size, type, attendance, dates, etc.) for use by other agents. Layout Agent: Ensures objects (stages, food vendors, activity zones, etc.) are well-placed, maximizing visibility, accessibility, and attendee engagement. Path Intelligence Agent: Optimizes paths and flow, preventing congestion and ensuring safe, logical movement across the venue. Validation Agent: Runs compliance checks against spacing, obstacles, amenities, and safety standards, surfacing critical issues. Safety Agent: Models safety risks (e.g., crowding hotspots, evacuation paths) and provides guidance for mitigating hazards. Crowd Agent: Simulates attendee behavior in real-time, predicting flow density and congestion zones. Optimization Agent: Continuously fine-tunes layout and flow for efficiency, balancing engagement, safety, and resource allocation. 2. Scenario and Pre-Risk Heatmap A floorplan for a hypothetical product exhibition hall was constructed. The scenario contains 300 attendees (starting locations indicated as blue dots). A Collision Free Speed Model simulation was generated via JuPedSim, such that the attendees independently navigate toward a central location with a single entrance (indicated in red), simulating a high anticipation event such as a new product showcase. Fig. 1: Floorplan of hypothetical product exhibition hall. Based on this floorplan, a heatmap representing relative risk of overcrowding was constructed by making an LLM API call feeding in contextual information along with positioning. Based on the contextual information as seen in the figure below, the LLM heuristically assigns relative risk levels. For example, an event that has already concluded is assigned lower relative risk, while the product showcase and associated bottlenecks as well as event registration are assigned higher relative risk. Fig. 2: Risk Heatmap 3. Single Agent Alerting A Proximal Policy Gradient (PPO) reinforcement learning model was trained to monitor the crowd risk level and issue binary alerts (i.e. {Alert, No Alert}) for a single agent in the crowd. The crowd risk levels were calculated using a combination of both the heuristic risk heatmap (Fig. 2) and real-time crowd dynamics data (e.g. crowd density and velocity entropy). The agent’s response to an alert was simulated as a reduction of the agent’s desired speed by some uniformly random factor. The reward function was configured to penalize elapsed time as well as putting the alerted agent in a location with a high risk score (e.g. heatmap-adjusted velocity entropy beyond a certain threshold). The model learns to alert the agent (denoted as a gold star below) to stay back behind the main crowd and away from the bottleneck, whereas without alerts, the agent is directly drawn into the bulk of the crowd. BEFORE (NO ALERTING) - (Video) AFTER (WITH SINGLE-AGENT ALERTING USING TRAINED PPO RL MODEL) - (Video) Fig. 3: Crowd simulation in hypothetical product exhibition hall without and with model alerting Note that the trained model learns to keep the agent away from the crowd situation. 4. Group Alerting The single-agent alerting model was extended to a PPO model that issues alerts to everyone in the crowd. Alert levels were enhanced from a binary {Alert, No Alert} system to a three-tiered multi-discrete alerting system with level-specific agent speed multipliers. The reward function incentivizes successful pedestrian exits and risk-delta improvements, and penalizes for high risk maxima, elapsed time and blanket alerting. The RL model’s actions (alert levels issued to each pedestrian in the simulation) are modulated by two overlays: firstly, a Graph Attention Network was added to capture local interaction topology via heuristic attention with distance kernels; secondly, Hamilton-Jacobi Backward Reachable Tubes are used to determine and avoid dynamically evolving hazard regions as captured by the risk grid, protecting “fast-lane/near-exit” agents from alert escalations (thus impeding evacuation). The PPO model was trained over a vectorized simulation environment across 6 parallel CPUs. BEFORE (WITH NO ALERTS, THE ENTIRE CROWD RUSHES TOWARD THE DESTINATION) - (Video) AFTER (AGENT-SPECIFIC ALERTING WITH MULTIPLE LEVELS OF SEVERITY ALLOWS FOR GREATER DISPERSION OF THE MOVING CROWD) - (Video) Fig. 3: Crowd simulation in hypothetical product exhibition hall, including randomly placed “bad agents” that do not reduce speed in response to alerts. Links: Pre-Event Heuristics and Simulation Demo Video Single Agent ("FPS") Alerting Reinforcement Learning Model Group Alerting Reinforcement Learning Model Demo videos on office egress scenario: (No Alerts) (Alerting Using Trained PPO Model) (Trained Model) Full Codebase