We present a pilot study on the implementation of a software, based on declarative knowledge representation and logic-based automated planning, which assists the management of severe Chemical hazard events, e.g. re or emissions that may require the evacuation of the area surrounding the a ected Chemical plant. We model the geography, the road network and the population of the chemical plant and the surroundings by weighted, labeled graphs, which are updated as the hazardous situation develops. Intervening factors, e.g. the spread of toxic in the air, are represented in the graph in terms of their e ects. Risk for the resident population and possible evacuation plans are evaluated and re-evaluated as the accident develops. Also evacuation plans are contingent and as conditions change may be re-evaluated from scratch; moreover, they may involve complex coordinated actions among the rescue units. Both the evaluation of the emergency scenario and the evacuation planning phases have been prototyped by means of an Answer Set Programming planner.
We present a pilot study on the implementation of the central component of a safety tool for risk evaluation and evacuation planning to be deployed in response to re or emissions due to accidents at a large Chemical plant5.
We have created an abstract model of the geography of the chemical plant and of its surroundings by a weighted, labeled graphs, which is updated as the hazardous situation unfolds. Evacuation from the surroundings of the plant is evaluated and planned wrt. to the graph representation. Intervening factors, e.g. the spread of toxic in the air, are represented in the graph in terms of their e ects. Evacuation plans are contingent and as conditions change may be re-evaluated 5 Due to legal reasons, at this stage we must omit the name of the plant and of the residential area for which the tool has been developed. from scratch. Both the computation of the likely e ects and the planning phase have been prototyped by an Answer Set Programming planner.
Answer Set Programming (ASP) [
ASP is the language by which we represent all types of knowledge required
to address this scenario: declarative knowledge about the surroundings,
procedural knowledge about actions (escape actions, take cover actions and so on),
contingencies, and planning as a domain-independent strategy to solve a motion
problem. In this sense, our approach is in the same vein as the pioneer work
of Zepeda and Sol [
The drafting of evacuation plans in emergency requires nding a sequence of actions that lead from an initial state, representing a risk scenario, to a goal objective, representing a situation where the entire population is rescued (or generally safe). Unlike the planning scenarios that are traditionally considered in the Arti cial Intelligence literature, evacuation plans specify administrative/security policies which can be hard to formulate6|even informally and often hard to execute even in small scenarios, i.e., those where the number of subjects, the spacial dimension and the time-scale are reduced.
Another important di erence is the value to give to the 'do nothing' action. While in AI planning the so-called nop action is there mostly for padding xedlength plans, in our applicative scenario they have a precise meaning which must be re-evaluated constantly: in case of chemical hazard, staying inside the building and limiting air circulation could be safest option available. Therefore we can say that automated planning with AI techniques, which is the subject of this 6 See the norms regulating Save protezionecivile.gov.it/ & rescue in Italy from http://www. paper, is only one dimension of the inherent complexity of emergency evacuation management.
The other key element to the formalization of evacuation plans is the representation of the area, with a dynamic description of elements such as i) source and type of hazard ii) risk di usion maps, which are speci c to the type of risk, i.e., iii) number and localization of the population that needs to be evacuated iv) transport means and their level of mobilization v) Accident & Emergency (A & E) services with trained personnel and specialized equipment. For know risks, normally associated to Chemical/energy plants, the complexity of the task is essentially decreased by the availability of pre-compiled maps, which can specify the following two types.
First, during the emergency the danger areas extend (or contract) following the evolution of the accident, moreover such expansion/contraction is not easily characterized by simple circumferences around the site of the accident (consider, e.g., liquid chemicals in rivers, or re under constant-direction winds). Normally, risk di usion maps create a three-level partition of the areas in i) impact, i.e., areas close to the epicenter of the disaster, with high likelihood of lethality, ii) damage, normally external to the former, where lack of protection would cause irreversible damage to those who are contaminated, especially children and old people, and iii) attention, where damage is possible but not irreversible, in any case requiring medical treatment and possibly causing unrest in the population.
Second, so-called safe areas and their features. These areas are further detailed in i) waiting areas, ii) concentration areas for the rescuers and iii) recovery areas, which are safe places where the refugees will end up as a result of the evacuation. 3
Representation of the geography and of the escape scenarios Two key aspects of the knowledge representation needed for this planning instance are the representation of the geography, namely roads and rivers, and of the level of risk assigned to areas by the domain experts. These information are synthesized by the risk graph, which is reported in Figure 1 for the rst instance of problem we considered.
For comparison, we report in Figure 1 one of the annotated maps, in the standard format for Geographical Information Systems (GIS) that have been used to compile the graph in Figure 1. The twenty-one relevant area (called waypoints) identi ed by the domain experts (in this case, Fire patrol senior o cers) are connected by 36 relevant routes.
As it can be noticed in Figure 1, domain experts have assigned each waypoints to one of 5 levels of risk for the population, according to the following standard risk scale.
{ RiskLevel = 1: recovery area, destination for evacuation plans;
{ RiskLevel = 2: low-risk area, close to recovery areas and far from the risk
areas;
{ RiskLevel = 3: average-risk area, far from recovery areas but su ciently
close to the risk areas;
{ RiskLevel = 4: high-risk area, close to high-risk areas, thus very far from
recovery areas;
{ RiskLevel = 5: high-risk area, the starting point for evacuation plans.
This subsection describes the ASP predicate de nitions that have been
developed to capture the speci c aspects of the evacuation planner. The following
description does not cover the general part of the evacuation planner, which has
been adopted as is from the ASP translation of the action description languages
L0 and L1 developed in [
It should be added that some simplifying assumptions have been embodied directly in the ASP representation of the planning instance. These assumptions are taken from the pre-compiled evacuation plan now in use, e.g., that for all evacuation actions there are some vehicles, typically buses, available to transport people to recovery areas. Another simpli cation is in the consideration of groups of evacuees, as opposed to single individuals. The structure of the graph in Figure 1 is embedded into the ASP program by means of the waypoint relation: group(N ame): where Name is assigned to thus-formed groups. The cardinality is not speci ed but as noted above we assume than one vehicle can evacuate a group. The last type of extensional predicate is for representing the position of the groups on the map, by this type of facts:
waypoint(N ame; RiskLevel): where variables Name and RiskLevel indicate the vertex of the graph and its assigned risk level. Communication routes, i.e., the edges of the graph are described by facts of this type: route(P lace1; P lace2; Length; Criticality): (2) where variables Place1 and Place2 indicate the two areas that are connected, Length captures geographical distance and Criticality is a parameter representing the danger assigned to the usage of the given connection. Another important information is the representation of the evacuee groups: (1) (3) holds(position(Group; P lace); 0): (4)
Notice how relation position(Group; P lace) is rei ed into a uent; variables Group and Place have their obvious meaning, whereas time-stamp 0 relates these facts to the initial state of the planning activity. Of course, we can have more than one group sitting on the same waiting area, as well as empty waiting areas. Finally, to describe actions where a certain areas become unreachable, i.e., a communication route has become nonviable (e.g., busy or disrupted or dangerous), we use these types of fact: waypoint blocked(W P 1): route blocked(W P 1; W P 2): (5)
The predicate described above are to be added to the domain description and changed often, to adapt to the changing scenario, especially the (possible) disruption of roads, to be acquired, in the full version of this planner, from real-time GIS information. 4
One of the most important problems to be solved in case of disasters is the draw and quick deployment of evacuation plans for the population. We describe a methodology based on knowledge representation and reasoning to formulate Evacuation Plans, using the intermediate graph representation and the DLV inferential engine.
Our evacuation planner considers the present situation, the type of danger, weather conditions, tra c or other modi cation of the zone to be evacuated, and formulates alternative evacuation plans to be face-validated on a case-bycase basis.
Studying the real case of the External Emergency Plan for a Re nery, we have implemented a planner able to generate appropriate evacuation plans, on the basis also of incomplete information derived from a possible Geographic Information System, supplying a representation of the scenario on the ground.
The results against a benchmark of 5 realistic emergency scenarios are
encouraging: computation times remain within few minutes and the generated
solutions were rated \excellent" by domain experts. From the point of view of
computational logic, these results are entirely satisfactory, and, in our opinion,
should become even more signi cant and widely applicable by the introduction
of two further formal devices. The rst device is the formal apparatus of ASP
programs with weak constraints developed by Leone et al. and implemented in
DLV. Even though there have been successful applications in literature, e.g. [
The second improvement would be a full model of context to be applied to
data, i.e., to redesign the data as to capture their contextual aspects, and have
the devices, e.g., local-cell emergency broadcasting, to selectively handle them,
along the lines of the methodology de ned Rauseo et al. [