Professional Documents
Culture Documents
°
c 2000 Kluwer Academic Publishers. Manufactured in The Netherlands.
Abstract. This paper describes an AI system for planning the first attack on a forest fire. This planning system
is based on two major techniques, case-based reasoning, and constraint reasoning, and is part of a decision support
system called CHARADE. CHARADE is aimed at supporting the user in the whole process of forest fire manage-
ment. The novelty of the proposed approach is mainly due to the use of a local similarity metric for case-based
reasoning and the integration with a constraint solver in charge of temporal reasoning.
Keywords: case-based reasoning, planning, local similarity metric, learning, temporal reasoning, constraint
satisfaction
is used mainly for managing temporal relations defined territory. So managing a forest fire can require several
on sub-plans or between actions in plans. Constraints centers to cooperate. Moreover, complex coordination
are also used in the identification of all the actions that problems can arise when resources from different or-
could be potentially applicable in a given emergency ganizations, such as forestry, police, and physicians,
scenario. The temporal dimension of the plan is man- are used.
aged with a quantitative approach based on temporal These are all general features of the fire domain. To
variables over continuous domains. Constraint reason- be more specific, the management of forest fires follows
ing also supports the resource allocation process for the a precise operational work-flow that is typical of each
chosen plan. We designed new and efficient temporal fire fighting organization. Here we focus on the work
reasoning procedures, in particular a fast incremental organization in a French regional center.
algorithm for updating temporal constraint networks.
The paper is organized as follows. First, we describe
the forest fire fighting domain, stressing its complex- 2.1. The Operational Context
ity. Then we illustrate the operational context where
the CHARADE system is used and the global system To understand how the user can be supported by the
architecture. In the following sections, we focus on CHARADE system in planning an intervention, a typ-
the interactive planning problem and on the proposed ical session of crisis management is illustrated.
solution. Section 3 describes the decision process re- The user of the system is a fireman working in a re-
lated to the interactive planning problem. Section 4 gional center in charge of supporting local fire fighting.
illustrates the plan representation adopted. A more de- His tools are a workstation, a dedicated line to acquire
tailed view of plan retrieval and plan reuse is given data from infrared sensors and meteo sensors, a radio,
in Section 5 and Section 6 respectively. The first de- a fax, a telephone and a printer.
scribes the local similarity metric introduced and the The CHARADE system is running on a workstation
second the efficient incremental algorithms for tempo- and comprises a geographic information system (GIS),
ral reasoning developed. We conclude by reviewing a a graphical simulator of the fire evolution, tools for ter-
few related works that share with our system both the ritorial, meteo and resource assessment, and a module
techniques (case-based reasoning) and the application for supporting the intervention planning and control.
domain (crisis response management). When a new fire is reported, the alarm is promptly
validated and the situation assessed by the user, who
possibly runs a mathematical fire spreading model. On
2. Planning for Fire Fighting the screen, the operator can look at the output of the fire
spreading model and access, through a graphical user
The complexity of the forest fire domain comes from interface, information on the graphical symbols shown
features that are typical of environmental problems. by the map. At the end of this phase, the operator has
Fire is a dynamic phenomenon whose evolution is acquired enough information for drawing on the map
determined by weather conditions, in particular wind a number of line sectors that subdivide the original fire
intensity and direction, by humidity, and by fuel type. front.
These variables usually change rapidly and sometimes This operation, called sectorization, involves crucial
in an unpredictable way. Moreover, relevant fire events and strategical decisions: on which fire front to attack
can happen on very different time and spatial scales, the fire; what specific technique to use on the selected
from seconds to days, from meters to kilometers, deter- fire fronts; how far from the fire epicenter to locate the
mining a large variety of world states. Data are always resources; what kind of supplies need particular atten-
incomplete and uncertain, and, in some cases, totally tion (for example inhabited houses, railway lines, etc).
absent. Once the sectors have been identified on the map, the
Operational constraints often impose quick deci- operator looks for a plan to fight the fire in each sec-
sions that drastically limit the possibility to build an tor. The plan may use air forces and/or ground means,
optimal plan. Sub-optimal solutions are often adopted and adopt a specific course of action. Searching in a
and are strongly biased by the past experience. database of past sector plans, the system retrieves a set
The management of forest fires, as in general en- of plans that in a similar context worked successfully.
vironmental emergencies, involves organizations that Then the retrieved plans are modified and adapted to
have decisional and operative centers distributed on the the current situation. These plans are evaluated by the
Case-Based Planning for Forest Fire Management 43
performed by the user or the planner: plan retrieval, or qualitative temporal constraints, while the planner
plan selection, plan reuse, plan refine, plan execution, checks the consistency of the whole intervention plan.
and plan recording. Let us describe them in detail: The plan applicability is checked up on problems that
can arise in the resource allocation phase.
Plan retrieval. The first task is performed by the plan-
ner that by looking into past intervention memory pro- Plan execution. The complexity derived from the
vides the user with a set of candidate solutions that highly dynamic and unpredictable domain of forest
were applied in similar circumstances. A good similar- fires makes plan verification very difficult by simula-
ity metric is crucial to provide an effective association ting plan execution. This task is done in the real world
between fire scenarios and firefighting plans. or is submitted to the evaluation of an expert.1
Plan selection. Among the proposed past solutions, Plan recording. At the end of the whole planning cy-
the user selects the candidate that better fits the current cle the user can decide whether to store the current case,
problem. Here many unstructured factors, too difficult both the fire scenario and the firefighting plan, in the
to be modeled, contribute to the human decision. case memory. Moreover, the planner itself filters out
some cases. That is done for efficiency and accuracy.
Plan reuse. Once a past solution is selected, the plan-
ner sets up a new plan reusing that past plan. Require- The next sections focus on plan retrieval, plan reuse,
ments derived by actions inside the plan need to be and plan refinement. For each step, it is shown how
combined with the constraints posted by the current case-based and temporal constraint reasoning tech-
situation. The result is a computer oriented represen- niques were extended to efficiently support forest fire
tation of a firefighting plan, as described in the next fighting.
Section.
4. Plan Representation
Plan revision. Past solutions do not always fit the
current problem and further adaptation is required to The system uses two different plan representations: one
develop an effective intervention plan. The revision for the plan that is developed in response to a given
task is accomplished by interleaving user and planner crisis, that is, the active plan, another for the plan stored
activities: the user provides new action specification in the plan’s memory. Here, we refer to the active plan
Case-Based Planning for Forest Fire Management 45
Figure 3. Part of the hierarchy in plan representation. Two examples of action net objects are shown; only the main attributes are reported.
and describe how we dealt with the critical requirement the plan it belongs to (an super) and to its sub com-
of making the plan representation sharable between the ponents (an subs), as shown in the example depicted
two main actors of the problem-solving process: the in Fig. 3.
system’s user and the planner. A plan component contains a set of domain depen-
The solution proposed in CHARADE rests basi- dent information, such as data resulting from the as-
cally on the following points. First, we used a rich sessment of the fire alarm, that are coded into the
representation for an intervention plan that points to scenario component of the action net structure. Dif-
data dynamically produced by the situation assess- ferent data are relevant at different abstraction levels.
ment component and to static, domain dependent So, for instance, the scenario at the action level contains
information, represented in the action library. This in- information specific to a fire fighting technique such as
formation can be browsed by the user by specific MMI the type and the number of resources to be used, the
components. Second, we adopted two different repre- types of actions that can or must be performed together
sentations for the plan temporal constraints, one user- to be effective (for instance spraying retardant in order
oriented, the other at use of the system, and provided to lower the fire intensity where squads will operate),
a mechanism for the automatic translation from one constraints that model the domain expertise on action
representation to the other. Moreover, we casted in a dimensioning (i.e., relations between the amount of a
single plan representation the possibility of reasoning specific resource and the amount of work that can be
following different firefighting schema that range from accomplished in a given time period). The scenario
planning the intervention for the global fire to planning at the sector plan level contains spatial information
the intervention for a single fire front, and gave the (sector length and orientation, accessibility, availabil-
possibility to switch from one scheme to the other. ity of water reservoirs near the sector, type of vegeta-
An active plan in CHARADE is represented as a hi- tion), fire’s physical parameters (fire intensity, flame
erarchy of plan components. The root represents the height, spreading direction and speed), information on
global intervention plan that can be composed of dif- the availability of resources located in the bases closest
ferent plans, one for each of the fire front sectors on to a sector.
which the user intends to act. The leaves correspond The temporal information relative to a plan compo-
to single actions. This representation gives a uniform nent is coded in a temporal constraint network asso-
description of the different abstraction levels at which ciated to its action net. The variables of this tem-
planning is performed, such as planning the interven- poral network are the start and the end time points of
tion on one or more fire front sectors or reasoning on the actions, of the sector and the global plans. The
a global plan that includes one or more (up to four) constraints are those representing the minimum dura-
sectors or checking the applicability conditions and di- tion of actions and those representing the interval re-
mensioning actions. lations between actions (for instance, the precedence
The basic element of the hierarchical representation relation). Moreover, the part-of relation between the
is the action net structure that contains references to sector plan and each one of its actions induces a pair of
46 Avesani, Perini and Ricci
temporal constraints between the sector’s start and the represented in this way with the exception of the in-
action’s start and respectively between the action’s end terval relations “before” and “overlaps” (and their
and the sector plan’s end (sectorstart − actionstart ≤ 0, inverses) the extent of which can be further spec-
actionend − sectorend ≤ 0). Analogous constraints ex- ified by a minimum and a maximum numerical
ist between the start and the end of the global plan parameter.
and the start and the end of the sector plans it is com- 2. The planner level: Here, time points are defined and
posed of. time constraints representing bounded differences
These constraints were modeled in terms of bounded between time points are implemented. The interval
difference between time points. This defines a con- relations are mapped to a set of metric constraints
straint satisfaction problem (CSP) on variables with representing bounded differences between the inter-
continuous domains [3], the so called Simple Temporal val endpoints. An example is given in Fig. 4.
constraint satisfaction Problems (STP).
Constraints on time points allow representing and Figure 5 shows a partial snapshot of the MMI inter-
reasoning about quantitative information, as discussed vention plan editor. This snapshot highlights how the
in Section 6, but cannot be easily grasped by the user. temporal structure of the plan is presented to the user.
Intervals and constraints between intervals are more An action is represented by a rectangular box with its
easily managed by the user. So the temporal dimension left side corresponding to the action’s earliest start time
of a plan is represented at two levels. and the right side to the latest start time. The inner area
filled in gray shows the action’s minimum duration.
1. The user level: Here, time intervals are con- Lines between boxes correspond to temporal relations
sidered and constraints between time intervals in between actions.
the form of single Allen’s interval relations [5] are The plan’s hierarchical structure is exploited for de-
handled. Mainly qualitative temporal relations are composing the temporal reasoning problem underlying
Figure 4. Examples of correspondence between interval relations and bounded difference constraints between the points representing the start
time and the end time of actions.
Figure 5. A snapshot of the plan editor. The plan’s temporal information is depicted.
Case-Based Planning for Forest Fire Management 47
the interactive planning process. That allows an effi- computed to obtain the relative distance from the probe.
ciency improvement, i.e., the temporal constraint net- A probe has a partial description of the fire case where
work of an action net contains temporal variables only the fire scenario is given while the definition of
and constraints that are relevant for the represented plan the plan strategy is still missing. The nearest case in the
element only. memory provides a candidate strategy to be used as a
Plan management functions supporting reasoning viable solution for the probe.
tasks specific to each different abstraction level are as- It is clear that the similarity metric is a key ele-
sociated with the action net structure. ment in the design of case-based system based on NN.
So, for instance, action allocation can be performed Usually, global metrics, which are generalizations of
only at the action level, insertion/deletion of actions and the Euclidean metric, are exploited. In this case, if
relations between actions can be performed only when cases x and y are described with n-dimensional vec-
working on sector plans, a sector plan can be deleted tors of numerical features (x1 , . . . , xn ), (y1 , . . . , yn )
only when working on the global fire plan. The user another vector of weights w = (w1 , . . . , wn ) is used
can easily shift from one level to the other. to balance the contributions
Pn of the features to the total
distance: (d(x, y) = ( i=1 wi |xi − yi |2 )1/2 ). Global
metrics have many limitations but essentially don’t
5. Plan Retrieval capture the correlation among features and the vari-
ability of feature relevance in different parts of the in-
The goal of the retrieval step is to find a past firefighting put space. For that reason, local metrics have been
plan to be used as a starting point for a new planning introduced [7].
phase. This provides the user with an example of a A local metric is a metric that depends on the point
possibly applicable strategy. in the inputP space from which the distance is taken
This goal can be achieved by considering the re- (d(x, y) = ( i=1n
wi (x)|xi − yi |2 )1/2 ). Local metrics
trieval step as a classification problem. Given a set of are context-sensitive, that is, similarity between cases
examples, each one belonging to one class in a given depends also on the absolute value of feature values.
finite set of classes, a classification problem consists of Using a local metric, it is possible to express condi-
assigning the right class to a new example whose class tional expressions like, “if the accessibility feature is
is unknown. To solve this problem, a set of features greater than 1000 meters, don’t use the orography fea-
(predictive) is used to derive the correct value of an ture to compute the similarity”.
unknown target feature, usually called class. Figure 6(a) illustrates the notion of a local similarity
In the forest firefighting domain, the classes are metric. A local metric is depicted as a box centered
the main strategies: ground attack, mixed attack and on the feature value and with width equal to the unit
aerial attack. The first strategy, ground attack, is ap- distance on the accessibility axis.
plied when aerial means can’t be deployed; in aerial
attack, the intervention can be managed only by aerial
means; in mixed attack, usually a firefighting plan in-
volves both kinds of means. The features describing
the fire scenario play a predictive role in finding a vi-
able strategy for a new case. The basic idea we use
to solve the classification problem and the initial plan
generation is to search in the memory for cases that are
similar to the one we want to solve and reuse one of
those plans that was used in similar situations. Simi-
larity is computed comparing the (predictive) features
of the current emergency scenario with those of the
stored one.
In other words, the retrieval step is based on the near-
est neighbor algorithm (NN), a well-known technique
to solve a classification problem [6].
Given a new probe emergency case, for each case
in the memory, a similarity-based distance function is Figure 6. Local weights and asymmetric metrics.
48 Avesani, Perini and Ricci
Case x#12 describes a case with accessibility 5.1. Local Weights and Asymmetric Metrics
value 800. Accessibility is the distance between the fire
front and the nearest road. The local weight value of The features used to describe a fire scenario can be
x#12 is 0.01; this implies that, along that feature, a unit categorical or real. In the first case, feature value varies
distance separates x#12 from cases with accessibility in a set of symbols; in the second case, it is normalized
value 700 or 900. to be a number in the closed unit interval [0, 1].
Let us suppose, as depicted in Fig. 6, that the Let F j be a generic feature space,
Qn i.e., a finite set
accessibility value is used to discriminate among three or the closed unit interval,
Q then j=1 F j denotes the
classes of intervention: ground attack, mixed attack and input space and x ∈ nj=1 F j a generic example. Each
aerial attack. For each strategy, a different set of actions space F j is endowed with a feature metric d j : F j ∪
can be combined to obtain a specific tactic. Usually, {?} × F j ∪ {?} → R≥0 , where ? is a special symbol
when it is easy to reach the fire front, (accessibility denoting an unknown value in F j :
<500 meters), a ground attack is chosen while when
the distance is greater than 1000 meters, only an aerial |x − y j | if x j , yi ∈ [0, 1]
j
attack can be effective. Our case memory and the asso-
0 if F j is categorical and
ciated local metric must incorporate this knowledge. If x j = yj
a new probe case has accessibility value 700, the nearest d j (x j , y j ) =
1
if F j is categorical and
neighbor found using the local similarity metric shown
x j 6= y j
in Fig. 6(a) is x#12. Therefore, the fire fighting plan
associated with x#12 is proposed, a plan based on a 0.5 if x j or y j is unknown
mixed attack strategy.
Using weights in distance computation poses an ad- A set of asymmetric weights w(x) for x is a 2 × n
ditional problem. Weights must be precisely chosen to matrix with values in [0, 1]. Let wkj (x), k = 0, 1 and
optimize system performance. We have proposed in [8] j = 1, . . . , n, be a generic element of w(x). We assume
a learning procedure for computing features’ relevance that w0j (x) = w1j (x) if F j is a set of symbols. Let y be
in a local metric. This procedure uses a reinforcement another point in the input space, if F j = [0, 1] then the
learning scheme and improves the retrieval accuracy of following notation will be adopted:
the initial local metric with all the weights equally set (
to the same value. w0j (x)d j (x j , y j ) if y j ≤ x j
Normally, both local and global metrics on real fea- w j (x) · d j (x j , y j ) =
w1j (x)d j (x j , y j ) otherwise
tures obey to the following property d(x, x + 1) =
d(x, x − 1). That means that the metric is invariant Conversely, when F j is a set of symbols w j (x) ·
with respect to the inversion of the direction of the d j (x j , y j ) = w0j (x)d j (x j , y j ) = w1j (x)δ j (x j , y j ).
displacement. Adapting weights in a metric with that Given a case base CB and a set of asymmetric
property may lead to some difficulties. For instance, weights for each example in CB, a local asymmetri-
given three cases x, y, and z, if a case x should be made cally weighted metric (LASM) is defined as follows:
closer to y and more distant from z, there are cases in
which one cannot attain that goal only changing the Y
n
local metric attached to x. d : CB × F j → R≥0
A greater flexibility is obtained if, for each feature j=1
a pair of weights is used, one for the “left” direction à !1/2 (1)
X
n
and one for the “right” direction [8]. In the example d(x, y) = w j (x) · d j (x j , y j )2
quoted above, if the metric attached to x is modified j=1
in such a way that the space in the direction y − z is
compressed and in the direction z − y expanded, the An example of such a metric is illustrated in Fig. 6(c).
effect is that y comes closer to x and z goes farther from For sake of simplicity here, only a feature dimension is
x. This cannot be achieved with a normal weighted shown, that associated with the accessibility. The met-
metric. ric defined in case x#12 is defined by two weights w0
In the following section, we give a more formal ac- and w1 . w0 (w1 ) is used when computing the distance
count of local metrics, and we describe the learning between x#12 and another case with accessibility less
procedure used for learning weights. (greater) than 800.
Case-Based Planning for Forest Fire Management 49
5.2. The Learning Procedure To describe this process in a more formal way, let
CB = {x1 , . . . , xm } be a subset of cases where (|CB|
Let us now focus on the learning procedure we use to is normally taken as 10% of C). The prototypes xi are
compute weight values. It is based on a reinforcement selected at random but the percentage of prototypes
learning scheme and iteratively transforms the weights in a given class is equal to the percentage of cases in
attached to a set of cases in the case memory using that class that were found in the original sample C. A
feedback given by the evaluation of the classification learning step is a pair of functions R and P that map a
response of the NN classifier. set of weights w(xi ) for xi , the example xi and a testing
The target problem isQto find an approximation of the example y in a new system of weights R(w, xi , y) and
target function Plan : nj=1 F j → S, where S is the set P(w, xi , y) for xi . R, the reinforcement step is chosen
of firefighting strategies (classes).
Q The Plan function if the value of Plan on y is equal to the value of Plan
is known only on a sample C ⊂ nj=1 F j and we want on xi . If this is not the case P, the punishment step is
to learn its behavior on the whole input space. used. A learning procedure iterates that step adjusting
The objective of the learning task is twofold: acquir- an initially constant system of weights for CB, aiming
ing the weight values and at the same time reducing the to optimize classification accuracy. Let xi ∈ CB be the
size of the case base. Since naive NN algorithms have nearest neighbor of y then the definition of R and P is
a linear time complexity with respect to the size of the shown in Fig. 7.
case base, using a subset of the case base (prototypes) α ∈ [0, 1] and β ∈ [0, 1] are called the reinforce-
yields a improvement in time performance. ment and punishment rate respectively. Note that each
Given a selection of prototypes CB ⊂ C, where C learning step updates at most n parameters and it can
is the complete set of cases for which a correct plan be shown that it maintains the weights in [0, 1] [12].
is known, our goal is to find a system of weights for The punishment function is chosen in such a way that
CB that maximizes the accuracy of the nearest neigh- the punishment, i.e. the value used to decrease wikj , is
bor algorithm, that is, E[Plan(nn(x)) = Plan(x)] is equal to βwikj d j (xi j , y j ) for 0 ≤ wikj ≤ 1/2, and it is
maximum,Q where nn(x) ∈ CB is the nearest neighbor equal to β(1 − wikj )d j (xi j , y j ) for 1/2 ≤ wikj ≤ 1 . This
of x ∈ nj=1 F j . behavior is needed to maintain 0 ≤ wikj ≤ 1 when pun-
To attain this goal, we proposed in [8] a training ishment is applied.
procedure that, given an example p in C but not in Finally, note that if xi j or y j are unknown values
CB, compares the value of Plan on p, which is known, and F j is a numeric feature, it is not possible to deter-
with the value of Plan on the nearest neighbor of p in mine whether y j < xi j so the maps R and P choose
CB. If the two values are equal, then the prediction is randomly which weight, among wi1j and wi0j , to update.
correct, and the distance between the nearest neighbor Figure 8 shows the pseudo code of the learning pro-
and p is decreased; whereas, if the two values are not cedure that takes as input a learning step map, a discrete
equal, the distance between the nearest neighbor and p function Plan, a sample C, a case base CB, a system of
is increased (see also [9–11] for other applications of weights for CB, and terminates, returning a new system
this learning strategy). Our approach is similar to the of weights for the original case base.
competitive learning technique [11] or learning vector In the Learn-Weights procedure, four auxiliary pro-
quantization [9]. But, while in these approaches the cedures are used.
cases stored are moved in the input space, we instead
change the metrics attached to the stored cases, and
therefore, directional information has to be taken into • Initialize-Weights(w): initializes all weights equally
account. (uniform deformation of the Euclidean metric).
For instance, consider the example in Fig. 6(d). It • Generate-Probe(C\CB): cyclically extracts all the
is simple to show that if the 15th and the 18th cases elements in C\CB.
are given, to obtain the maximal accuracy storing only • Retrieve(y, CB): finds the nearest neighbor of y in
an additional case, an abstract case with accessibility CB measuring the distances between y and points in
value equal to 750 would be needed. But that fire sce- CB with the current local metric definition.
nario could not belong to the real past experience. In a • Exit-Condition: stops the loop when the accu-
real domain, it is not advisable to create these abstract racy on C\CB decreases two times consecutively
cases; they could even be impossible due to other do- in each pass through C\CB. Accuracy is computed
main constraints. as nr/(nr + np), where nr and np are the number of
50 Avesani, Perini and Ricci
using the retrieved solution to guide the generative Then we will describe the temporal reasoning tech-
problem solver while searching a solution rather than niques exploited by these functions.
directly modifying the retrieved one.
This approach called generative adaptation [14] is
not suitable for intervention planning in the firefighting 6.1. System Functions
domain that lacks a generative problem solver.
In CHARADE, the adaptation task is performed in This section gives a description of the main functions
two steps: that can be called by the user during plan revision,
pointing out the specific kind of temporal reasoning
computations underlying each function call. An ex-
Plan reuse. As described in Section 4, the retrieved tended description of these functions can be found in
plan includes a set of constraints on time and resource [18, 19].
variables that needs to be reinstantiated in the con-
straint network associated with the active plan. These Inserting new actions or interval relations between ac-
constraints are automatically propagated and checked tions in a sector plan. Trying to insert a new ac-
against other descriptive variables of the current situ- tion in a sector plan requires updating the underlying
ation. For instance, if the candidate plan contains the temporal network by adding two new variables (the
action of spraying water by employing two squads, start and end variables of the new action) and at least
and, in the current emergency, the situation assessment three new constraints (the minimum action duration
calls for a greater amount of water to be sprayed, the and the two constraints induced by the part-of rela-
number of squads associated with the action will be tion). Analogously inserting an interval relation be-
recomputed. tween two actions requires adding the corresponding
distance constraints between the start/end variables of
Plan revision. A further adaptation phase is done the two actions. These modifications can cause a viola-
through an interactive process of editing activities per- tion of the sector deadline defined during the situation
formed by the user, like inserting/deleting actions, assessment phase, so the consistency of the updated
modifying action durations or temporal relations be- temporal constraint network associated to the sector
tween parts of the plan, and allocating resources. At plan must be checked. If no inconsistency is produced,
this level, plan adaptation exploits constraint reason- the action net structure representing the sector plan
ing techniques for modeling the constraints on actions is updated and the earliest and latest times of the vari-
and checking their consistency. In particular, this pa- ables of the sector temporal network are recomputed.
per focuses on the temporal constraints of a plan. For For instance, if the restored plan corresponds to the
example, a plan for controlling the fire on a given sector plan in Fig. 9(a), the current flame height at the head
can be impractical if the time required to perform it is sector makes a direct ground attack hard to perform,
greater than the deadline posed by the fire propagation and building a fire line would be advisable; the user
toward that sector. decides to insert the action building a fire line before
the action direct ground attack. This can cause a dead-
The exploitation of constraint satisfaction tech- line violation, as shown in Fig. 9(b), that can be man-
niques to adapt the candidate solution in the plan reuse aged by either attempting to delay the plan deadline or
and in the plan revision steps is becoming more and shortening the action duration allocating more squads
more common [16, 17]. on the actions.
In the following subsections, we focus our attention
on plan revision, a critical step in CHARADE since it Deleting actions or interval relations of a sector plan.
rests on an interactive process. Deleting an action or an interval relation between two
First, we describe the system functions for interac- actions corresponds to removing the associated con-
tive plan revision. Each function has been carefully straints (and the variables that are no longer linked)
designed in order to minimize computational cost. For from the sector temporal constraint network. This can-
instance, efficiency improvement has been obtained not produce inconsistency. The feasible time intervals
localizing constraint reasoning to components of the of the network variables can increase. So the variables
constraint network according to the plan hierarchical domain must be set to their default domains and the
structure described in Section 4. feasible times recomputed.
52 Avesani, Perini and Ricci
Figure 9. A plan, described as an action graph. The nodes, represented by gray boxes, correspond to actions; the edges, labeled by ovals, to
the temporal relations between actions. (a) The plan restored in the time window between 14 : 30 and 16 : 30. Here each action is described by
two overlapped rectangles: one, external, defined by the earliest start time and the latest end time of the action; the internal one, filled in gray, by
the earliest start time and the earliest end time. (b) Trying to insert a new action in plan (a). (c) Removing the temporal relation before between
two actions of plan (a).
For instance, deleting the temporal relation before Automatic allocation on the sector plan. The system
between the actions Retardant by helicopter and Water can be required to compute resource allocation on the
by plane increases the feasible interval of the start time sector plan. Some of the actions may be already al-
of Water by plane. This yields a greater flexibility of located (by manual allocation). The Allocator heuris-
these actions’ execution with respect to the global plan tically orders the actions to be allocated and sched-
deadline (Fig. 9(c)). ules them, setting the action start time to the earliest
possible time that is updated according to the already
Modifying the duration of an action. Modifying the scheduled actions. This corresponds to computing a
action minimum duration corresponds to a change solution (or completing a partial one) for the sector
of the relative constraint. Weakening the constraint temporal network.
(i.e., reducing the minimum duration) doesn’t produce
inconsistency; the previous feasible values of an ac-
tion’s start and end are still good values. These inter- 6.2. Temporal Reasoning
vals must be updated, computing new consistent val-
ues. When trying to increase the minimum duration, the As pointed out in the above discussion, the addition
consistency has to be checked against sector deadline (or deletion) of constraints to (or from) the constraint
violation. network associated with the current plan during the
Case-Based Planning for Forest Fire Management 53
Figure 11. An example of the typical temporal problem that can be represented with simple negative temporal networks.
reasoner computes the minimal network of the sector vertices are “collapsed” into a single vertex forming a
temporal network before starting automatic allocation metavertex. Two vertices x, y of a distance graph G
and supports the process of building a solution by up- which does not contain negative cycles are equivalent
dating the feasible values to the set of values that can (written as x ∼ y) if and only if there exists in G a cycle
complete a current partial solution. connecting x and y whose length is zero.4
Since G ∗ is acyclic, its shortest paths can be com-
puted by using a shortest-paths algorithm for DAGs
6.2.2. Incremental Algorithms for STP− . We now
such as DAG-Shortest-Paths given in [25]. DAG-
introduce a subclass of STP, called STP− , consisting
Shortest-Paths takes O(|E|) time and space, and thus
of the STP-constraints of the form x j − xi ≤ a, where
it is more efficient than the Bellman-Ford algorithm
a ≤ 0. STP− has an appealing feature: the distance
requiring O(|V ||E|) time. Insertion (or deletion) of
graph of a set of STP− -constraints can be incrementally
constraints is performed by checking possible modifi-
managed through efficient algorithms.
cations (inconsistency, expansion) of the G ∗ cycles and
Figure 11 shows an example of a plan whose tempo-
eventually updating the topological order of G ∗ ’s ver-
ral dimension can be represented as a simple temporal
tices before running DAG-Shortest-Paths algorithms
negative constraint problem (see also [18]). Actions are
that updates the graph’s shortest-paths.
represented with temporal intervals that are constrained
An experimental comparison of the proposed al-
to be within a time span given by two temporal vari-
gorithms with the (non-incremental) BF algorithm
ables, the source S and the end F. Constraints on the
showed that drastic CPU-time reductions can be
minimal duration can be expressed. Even if we cannot
obtained [24].
express explicitly a constraint on the maximal duration
of an action, we can place deadlines. The expressive-
ness of STP− is discussed more deeply in [24]. 7. Related Works in Crisis Management
Here we point out the ideas exploited in the incre-
mental updating algorithms. A detailed description is Crisis management poses challenging issues to be in-
given in [24]. The dynamic management of STP− - vestigated by AI researchers. One of the first attempts
constraints is based on a special data structure that we to approach the problem of the real-time management
called the “distance metagraph.” The distance meta- of forest fires is the Phoenix project [28]. Phoenix ad-
graph is derived from the distance graph of a given dresses two research objectives: the design of complete
STP (or STP− ) whose vertices are maintained topo- autonomous agents and the constraints that a complex
logically ordered upon addition/deletion of constraints. dynamic environment places on the design of intelli-
This data structure allows the applicability of known gent agents.
directed acyclic graph (DAG) shortest-path algorithms Several works in crisis management are reported
the extension of to the distance graph of a STP− . in the review conducted in the Esprit Project NOW
The distance metagraph of a distance graph, G ∗ , is [29]. Moreover, the interest in these research topics
a directed acyclic graph in which sets of equivalent was promoted also by Darpa projects [30, 31], where
Case-Based Planning for Forest Fire Management 55
Italsoft (Resource Allocator), Thomson CSF/LER (car- 12. F. Ricci and P. Avesani, “Nearest neighbor classification with a
tographic module). local asymmetrically weighted metric,” IRST Technical Report
#9602-01, 1996.
13. F. Ricci and P. Avesani, “Data compression and local metrics
Notes for nearest neighbor classification,” in IEEE Transactions on
Pattern Analysis and Machine Intelligence, vol. 21, no. 4, April
1. In the context of training, where the system is currently used, 1999.
the global evaluation of an intervention plan, as designed by the 14. M. Veloso, H. Munõz-Avila, and R. Bergmann, “Case-based
trainee according to the plan representation above, is done by the planning: Selected methods and systems,” AI Communications,
expert. vol. 9, no. 3, pp. 128–137, 1996.
2. The minimal network representation contains the network infor- 15. D. Leake, A. Kinney, and D. Wilson, “Acquiring case adap-
mation in the most explicit form; i.e., all the possible values of tation knowledge: A hybrid approach,” in Proceedings of the
the distance between the variables x, y are directly available from Thirteenth National Conference on Artificial Intelligence, AAAI
the label of the edge that connects x, y and, considering the rest Press: Menlo Park, CA, 1996.
of the network, does not produce any further reduction. 16. L. Purvis and P. Pu, “Adaptation using constraint satisfaction
3. E t = {(y, x) ∈ V × V | (x, y) ∈ E} and the weight of an edge techniques,” in International Conference on Case-Based Rea-
(y, x) in G t is the same as the weight of (x, y) in G. soning (ICCBR-95), Sesimbra, Portugal, October 23–26, 1995.
4. “∼” is an equivalence relation and if two vertices are equivalent, 17. W. Wilke, B. Smith, and P. Cunningham, “Using configuration
then they have the same s.p. distance from the source vertex of techniques for adaptation,” in Case-Based Reasoning Technol-
the graph. The graph G ∗ = (V ∗ , E ∗ ) is called the canonical ogy from Foundations to Applications, Springer-Verlag, 1998.
metagraph of G = (V, E). If G contains only loops whose 18. A. Perini and F. Ricci, “An interactive planning architecture,” in
length is zero, then the canonical metagraph is also known as the Third European Workshop on Planning, Assisi, Italy, Septem-
component graph. The vertices of the component graph are the ber 27–29, 1995. New Directions in AI Planning, IOS Press.
strong connected components of G [10]. 19. P. Avesani, A. Perini, and F. Ricci, “The intervention plan-
ning subsystem,” Technical Report, IRST, 1995. CHARADE
Restricted Report #50B.
References 20. C.E. Bell and A. Tate, “Use and justification of algorithms for
managing temporal knowledge in o-plan,” Tec. Rep. AIAI-TR-6,
1. Drew V. McDermott, “The current state of AI planning research,” AIAI University of Edinburgh, Edinburgh, U.K., 1985.
in Proceedings of the Seventh International Conference on In- 21. R. Cervoni, A. Cesta, and A. Oddi, “Managing dynamic tempo-
dustrial and Engineering Applications of AI and Expert Systems, ral constraint networks,” in Proceedings of the Second Interna-
1994. tional Conference on Artificial Intelligence Planning Systems,
2. K. Kolodner, Case-Based Reasoning, Morgan Kaufmann: San edited by K. Hammond, 1994, pp. 196–201.
Mateo, CA, 1993. 22. A. Gerevini, A. Perini, and F. Ricci, “Incremental algo-
3. R. Dechter, I. Meiri, and J. Pearl, “Temporal Constraint Net- rithms for managing temporal constraints,” in Proceedings
works,” Artificial Intelligence, 49, 1991. of the Eighth IEEE International Conference on Tools with
4. V. Normand, “Task modelling in the design and implementation Artificial Intelligence, Toulouse, France, November 16–19,
of interactive systems,” in Proceedings of the Fifth Conference 1996.
on the Engineering of Man-Machine Interfaces, Lyon, France, 23. U. Montanari, “Networks of constraints: Fundamental prop-
1993. erties and applications to picture processing,” Inf. Sci., vol. 7,
5. J.F. Allen, “Maintaining knowledge about temporal intervals,” pp. 95–132, 1974.
Communications of ACM, vol. 26, no. 11, pp. 832–843, 1983. 24. A. Gerevini, A. Perini, and F. Ricci, “Incremental algorithms
6. B.V. Dasarathy, Nearest neighbor (NN) norms: NN pattern for managing temporal constraints,” Tec. Rep. IRST-9605-07,
classification techniques, IEEE Computer Society Press: Los IRST, 1996.
Alamitos, CA, 1991. 25. T.H. Cormen, C.E. Leiserson, and R.L. Rivest, Introduction to
7. C. Atkenson, A. Moore, and S. Schaal, “Locally weighted learn- Algorithms, The MIT Press, 1992.
ing,” AI Review Journal, vol. 11, pp. 11–73, 1997. 26. E. Davis, “Constraint propagation with interval labels,” Artificial
8. F. Ricci and P. Avesani, “Learning a local similarity metric for Intelligence, vol. 32, pp. 281–331, 1987.
case-based reasoning,” in Proceedings of International Confer- 27. A.K. Mackworth, “Consistency in network of relations,” Artifi-
ence on Case-Based Reasoning, Sesimbra, Portugal, 1995. cial Intelligence, vol. 8, pp. 99–118, 1977.
9. T. Kohonen, “The self-organizing map,” in Proceedings of the 28. P. Cohen, M. Greenberg, D. Hart, and A. Howe, “Trial by fire:
IEEE, 1990, vol. 78, no. 9, pp. 1464–1480. Understanding the design requirements for agents in complex
10. D. Aha and R. Goldstone, “Concept learning and flexible weight- environments,” The AI Magazine, vol. 10, no. 3, pp. 32–48,
ing,” in Proceedings of the Fourteenth Annual Conference of the 1989.
Cognitive Science Society, Lawrence Earlbaum, Bloomington, 29. G. Toffoli, “Dissemination and exploitation of European RTD
IN, 1992, pp. 534–539. results in the area of the EDSS-related technologies”, The Esprit-
11. L. Xu, A. Krzyzak, and E. Oja, “Rival penalized competitive NOW Newsletter No. 1, May 1998.
learning for cluster analysis, RBF net, and curve detection,” 30. N. Fowler, S.E. Cross, and C. Owens, “The ARPA-Rome
IEEE Transaction on Neural Networks, vol. 4, no. 4, pp. 636– knowledge-based planning and scheduling initiative,” IEEE
649, 1993. Expert, vol. 10, no. 1, February 1995.
Case-Based Planning for Forest Fire Management 57
31. P. Cohen, R. Schrag, E. Jones, A. Pease, A. Lin, B. Starr, Lab working in Intelligent Tutoring Systems. Since 1989 he has been
D. Gunning, and M. Burke, “The DARPA high-performance a research scientist in the Artificial Intelligence Department at IRST,
knowledge bases project,” The AI Magazine, vol. 19, no. 4, working on AI applications in the domain of decision support sys-
pp. 25–49, 1998. tems and emergency planning problems. Currently he is involved
32. M. desJardins and M. Wolverton, “Coordinating planning ac- in a project aimed at developing a CBR tool that enables fine tun-
tivity and information flow in a distributed planning system,” ing of similarity metrics, incremental retrieval and partial matching
in AAAI Fall Symposium on Distributed, Continual Planning, between structured data. His research interests include Case-Based
AAAI Press, 1998. Reasoning, Machine Learning, and Planning. He is a member of
33. M. Veloso, “Towards mixed-initiative rationale-supported plan- AAAI and AIIA.
ning,” in Advanced Planning Technology, edited by A. Tate,
AAAI Press, pp. 277–282, May 1996,
34. M. Veloso, A. Mulvehill, and M. Cox, “Rationale-supported
mixed-initiative case-based planning,” in Proceedings of IAAI-
97, Innovative Applications of Artificial Intelligence, Provi-
dence, Rhode Island, July 1997, pp. 1072–1077.
35. M. Goodman, “CBR in battle planning,” in Proceedings of
the Second Workshop on Case-Based Reasoning, pp. 264–269,
1989.
36. D. Leake, A. Kinney, and D. Wilson, “A case study of case-
based CBR,” in Proceedings of the Second International Con-
ference on Case-Based Reasoning, Springer Verlag: Berlin,
1997. Anna Perini received her doctoral degree in physics from the
37. M. Gervasio, W. Iba, Langley, P., “Case-based seeding for University of Trento in 1981. She worked in the area of statistical
an interactive crisis response assistant,” in AAAI-98 Pro- mechanics and computer simulation applied to solid state physics
ceedings, Workshop on Case-Based Reasoning Integrations, problems, with a fellowship from the Istituto per la Ricerca Scien-
1998. tifica e Tecnologica, Trento (itc-IRST), Italy. Since 1985 she has
38. A. Aamodt and E. Plaza, “Case-based reasoning: Foundational been a Research Scientist in the Artificial Intelligence Department
issues, methodological variations, and system approaches,” AI at IRST, working on the application of AI techniques to environ-
Communications, vol. 7, no. 1, pp. 39–59, 1997. mental decision support systems devoted to agricultural and to forest
39. F. Ricci, P. Avesani, A. Perini, and M. Stocchero, “Approaches fire management problems. Her research interests include constraint
and techniques for AI planning,” CHARADE Technical Report satisfaction problems and temporal reasoning, case-based reasoning
RR51, 1993. and planning.
40. F. Ricci, P. Avesani, and A. Perini, “Cases on fire: Applying
CBR to emergency management,” The New Review of Applied
Expert Systems, 1999, to appear.
41. R. Short and K. Fukunaga, “A new nearest neighbour distance
measure,” in Proceeding of the Fifth IEEE International Confer-
ence on Pattern Recognition, Miami Beach, FL, 1980, pp. 81–
86.
42. C. Stanfill and D. Waltz, “Toward memory-based reasoning,”
Communications of ACM, vol. 29, pp. 1213–1229, 1986.
43. S. Terral, “Fighting talk,” GIS Europe, vol. 5, no. 6, pp. 46–48,
1996.
44. I. Watson, Applying Case-Based Reasoning: Techniques for En-
terprise Systems, Morgan Kaufmann, 1997. Francesco Ricci was born in Venice, Italy, in 1959. He received
the doctoral degree in Mathematics from the University of Padova,
in 1983. From 1984 to 1986, he was Ph.D. student in Mathematics
at Scuola Normale Superiore, Pisa, Italy. From 1986 to 1987, he
was with Enichem, S.P.A., Milan, Italy, where he engaged in de-
veloping expert systems and office automation applications. Since
1988 he has been a researcher at Istituto per la Ricerca Scien-
tifica e Tecnologica, Trento. His current research interests include
constraint satisfaction problems, machine learning and case-based
reasoning. Dr. Ricci has been on the programme committee of in-
ternational conferences and is co-chair of EWCBR-2000 in Trento.
Paolo Avesani received his doctoral degree in Information Science Dr. Ricci received the best degree award from University of Padova in
from the University of Milan in 1988. He spent two years in the Dida 1984.