# theoretical_study_on_multiobjective_heuristic_search__632e4ee3.pdf

Theoretical Study on Multi-objective Heuristic Search

Shawn Skyler1 , Shahaf Shperberg1 , Dor Atzmon2 , Ariel Felner1 , Oren Salzman3 , Shao-Hung Chan4 , Han Zhang4 , Sven Keonig4 , William Yeoh5 , Carlos Hernandez Ulloa6

1Ben-Gurion University of the Negev 2Bar-Ilan University 3Technion Israel Institute of Technology 4University of Southern California 5Washington University in St. Louis 6Universidad San Sebasti an {shawn@post.bgu, shperbsh@bgu, dor.atzmon@biu, felner@bgu, osalzman@cs.technion}.ac.il, {shaohung, zhan645, skoenig}@usc.edu, wyeoh@wustl.edu, carlos.hernandez@uss.cl

This paper provides a theoretical study on Multi Objective Heuristic Search. We first classify states in the state space into must-expand, maybe-expand, and never-expand states and then transfer these definitions to nodes in the search tree. We then formalize a framework that generalizes A* to Multi Objective Search. We study different ways to order nodes under this framework and its relation to traditional tie-breaking policies and provide theoretical findings. Finally, we study and empirically compare different ordering functions.

1 Introduction

In Single-Objective Search (SOS), the task is to find a leastcost path between two states. Multi-Objective Search (MOS) generalize SOS to the case where each edge in the graph is associated with multiple costs. Examples include pathfinding applications with multiple competing resources, such as time and fuel as well as many other applications [Bachmann et al., 2018; Fu et al., 2019; Fu et al., 2021]. In MOS, the task is to find a (cost-unique) Pareto-optimal frontier (POF), which is a set of undominated paths between start and goal, i.e., there is no other path with lower cost in all objectives. Various MOS algorithms were developed [Cl ımaco and Pascoal, 2012; Current and Marsh, 1993; Skriver, 2000; Tarapata, 2007; Ulungu and Teghem, 1991; Mandow and P erez-dela-Cruz, 2010], as well as Bi-Objective Search (BOS) algorithms for solving the two-dimensional MOS problems [Ahmadi et al., 2021; Skyler et al., 2022; Hern andez et al., 2023b]. A recent survey including open challenges has also appeared [Salzman et al., 2023]. A classic theory for SOS characterizes the set of states that any unidirectional search algorithm must expand to prove the optimality of the solution [Dechter and Pearl, 1985]. While investigating the NAMOA algorithm, Mandow and P erezde-la-Cruz (2010) (we denote that paper by MPC10) provided theoretical insights into MOS and identified must-expand and

never-expand nodes, where each node is associated with an undominated path to some state in the graph. In this paper we provide a theoretical study of MOS algorithms. First, we extend the work of MPC10 and present a comprehensive and unifying theory for the MOS setting. We introduce an alternative analysis that shifts the focus from nodes to states, providing a theory that classifies MOS states into must-expand, maybe-expand, and never-expand categories. We demonstrate that an algorithm sharing information among nodes corresponding to the same state need not expand all must-expand nodes as defined by MPC10 but has to expand all must-expand states, as per our definition. We then formalize an algorithmic framework called MOSA , which generalizes A [Hart et al., 1968] to MOS and encompasses various existing MOS algorithms. In MOS-A , a node with an undominated f-value is expanded in each iteration. In MOS, f-values lack a total order, and thus, multiple nodes with different undominated f-values may exist simultaneously. Therefore, an Ordering Function (OF) is introduced to impose a total order on undominated f-values. Additionally, a tie-breaking policy (TB) can be employed to resolve ties between nodes with the same f-values. We restate the must-expand and never-expand nodes identified by MPC10 in the context of MOS-A and characterize maybe-expand nodes. Moreover, we formally prove that the choice of OF does not impact which nodes are expanded during the search, while different TBs can result in different node expansions. However, OFs can influence the rate and order by which solutions are discovered, which can be crucial when aiming for quick solution discovery in an anytime manner. To this end, we introduce several OFs, analyze their impact on solution discovery, and empirically compare them on roadmaps from [DIMACS, 2006] for MOS with 2 4 objectives.

2 Definitions and Background A Multi-Objective Search graph G with d objectives is a tuple S, E, c , where S is the finite set of states, E S S is the finite set of edges, and c : E (R 0)d is a cost function in the form of a d-dimensional vector (one dimension per objective) of non-negative costs. A path π from s1 to sm

Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence (IJCAI-24)

is a sequence of states s1, s2, . . . , sm such that (si, si+1) E for all i {1, . . . , m 1}. Boldface font is used to represent d-dimensional vectors in the form v = (v1, v2, . . . , vd). For two vectors u and v, u+v = (u1 +v1, . . . , ud +vd). c(π) = Pm 1 i=1 c(si, si+1) is the cost of path π = s1, . . . , sm. Given two d-dimensional vectors u and v, we say that u weakly dominates v, denoted as u v, if ui vi, for every i {1, . . . , d}. Vector u (strictly) dominates v, denoted as u v, if u v and u = v (i.e., for at least one dimension i, ui < vi). If ui > vi for some i {1, . . . , d}, then we say that u undominates v (likewise v is undominated by u), and denote this by u v. For example, (4, 5) (4, 5) (they are in fact equal), (4, 5) (4, 7) (as 5 < 7) and (4, 7) (4, 5). Finally, we say that two vectors u and v are mutually undominated if both u v and v u. For example, (4, 8) and (5, 6) are mutually undominated. Similarly, we say that path π dominates (weakly dominates) path π , denoted as π π (π π ) if c(π) c(π ) (c(π) c(π )). A search instance is a tuple I = S, E, c, start, goal, h , where S, E, c is a MOS graph, start, goal S are the start and goal states, and h is a heuristic function. A solution is a path from start to goal. A solution is Pareto-optimal if it is undominated by any other solution. A Pareto-optimal solution set for I, denoted by POS, is the (maximum) set of Pareto-optimal solutions. Let POC = {c(π) | π POS} denote the set of path costs in POS. By definition, any two members of POC are mutually undominated. However, there can be two paths with equal costs in POS. We thus define a cost unique Pareto-optimal solution set, also called the Pareto-optimal frontier, denoted by POF, as a subset of POS such that no two paths have equal costs. That is, there is only one path for every value in POC. In the special case of SOS (where d = 1), POS includes all optimal-cost paths while POF includes only one optimal-cost path. Figure 1 shows different cost vectors of a BOS problem instance (where d = 2). The C points denote the POC. This means that there are no solution costs in zones A or B. Similarly, solutions in C are undominated by any solution in Zone D. We consider Single-Valued Heuristic functions which assign to each state a d-dimensional vector h = {h1, . . . , hd} where each value hi estimates the cost along dimension i. h is admissible if hi is a lower bound of all the costs of dimension i in any member of POC. Similarly, h is consistent iff (i) h(goal) = 0 and (ii) h(s) c(s, t) + h(t) for all (s, t) E. The strongest possible single-valued heuristic, which is commonly used, is the individual shortest path heuristic, also called (Single) Point heuristic [Goldin and Salzman, 2021; Hern andez et al., 2023b; Pulido et al., 2015; Skyler et al., 2022]. That is, for h(n), hi(n) is the costminimal path from n to goal using the i-th objective only. In Figure 1, S represents the value of the Point heuristic, and Zone H A includes all other admissible heuristics.

2.1 Classes of States in Single-Objective Search Following Dechter and Pearl (1985), let IAD denote problem instances that have an admissible heuristic, and let ICON IAD denote the instances that have a consistent heuristic. Admissible algorithms are algorithms that guarantee to return optimal solutions on all problem instances from IAD. Like-

Figure 1: Different zones in a BOS problem instance. Every point in the figure is a two dimensional cost from start to goal.

wise, a search algorithm is said to be DXBB [Eckerle et al., 2017] if it is deterministic and has only black-box access to the graph and the heuristic, i.e., it can only discover states, edges, and cost by continuously applying an expansion operator from either start or goal. Admissible SOS algorithms must accomplish two tasks: (i) find a solution of cost C and (ii) prove that the returned solution is optimal. A classic theory is that any admissible DXBB unidirectional heuristic search algorithm must expand all states s such that there exists a path U from start to s for which c(U) + h(s) < C , when running on an instance from ICON [Dechter and Pearl, 1985]. Such states are called Must-Expand States (MESs). To fulfill task (ii), algorithms must expand every MES s or they might miss an optimal solution that passes through s. In addition, if c(U)+h(s) > C for all paths U from start to s, then state s should never be expanded, as it cannot lead to an optimal solution and, thus, does not contribute to accomplishing either tasks such states are called Never-Expand States (NESs). Finally, depending on the tie-breaking policy (TB), some states s, where c(U)+h(s) = C , might be expanded to find an optimal solution (task (i)) and are called Maybe-Expand States (MBESs). In the next sections, we generalize this theory from SOS to MOS. We distinguish between states in the graph and nodes in the search tree. We dedicate a separate section to each, motivated by the inherent differences between the two entities, as demonstrated in Section 4.4. It is noteworthy that the theoretical study of MCP10 (while exploring the NAMOA algorithm) partially overlaps with our current work. Specifically, they defined the concepts of must-expand paths and never-expand paths (where paths in their notation align with nodes in our notation). As their definitions and corresponding proofs are pertinent to our discussion (Definitions 1,2 for states and Definitions 4,5 for nodes), we restate them here using our terminology for the sake of completeness.

3 Classes of States in Multi-Objective Search

In MOS, the two aforementioned tasks are modified as follows: (i) find all values of POC and (ii) prove that they are undominated by any other costs of solutions.

Definition 1 (Must-Expand States (MESs)). A state s is a MES iff there exists a Pareto-optimal path U from start to s such that p c(U) + h(s) for all path costs p POC.

We partition the set of MESs into: Optimality States and Completeness States; their union is the set of MESs. Definition 1.1 (Optimality States). A state s is an optimality

Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence (IJCAI-24)

state iff there exists a path U from start to s and p POC such that c(U) + h(s) p. Optimality states are expanded to prove that any path in a POF is Pareto-optimal, i.e., that there is no path π from start to goal that dominates some path π POF. Optimality states are analogous to MESs in SOS (with f < C ). Zone A in Figure 1 corresponds to costs of optimality states (each via a relevant path U with the corresponding f-value). In SOS, there is a total order between the cost of edges (and therefore between paths in the graph). However, in MOS, there is only a partial order because the costs of two edges may be mutually undominated. Thus, MOS algorithms must also consider states such that the costs of paths through them are mutually undominated with any cost in POC and prove that such costs cannot be added to POC. Such states are called completeness states. Definition 1.2 (Completeness States). A state s is a completeness state iff there exists a path U from start to s such that p POC: c(U) + h(s) p and p c(U) + h(s) Zone B corresponds to costs of completeness states (via a relevant path U). These states must be expanded to ensure that there are no more solution costs in POC (i.e., POC is complete). Completeness states have no analogy in SOS. Definition 2 (Never-Expand States (NESs)). A state s is a NES iff for every path U from start to s there exists p POC such that p c(U) + h(s). Zone D in Figure 1 corresponds to NESs. They are analogous to the Never-Expand States in SOS. Definition 3 (Maybe-Expand States (MBESs)). A state s is a MBES iff s is not a MES and there exists a path U from start to s and a cost p POC such that c(U) + h(s) = p. Zones C1 to C4 correspond to the costs of MBESs, analogous to states with f = C in SOS. At least one goal state from each Ci zone is expanded, but depending on the tie-breaking policy, other states with the same costs may be expanded too.

3.1 Theoretical Analysis We now prove that all MESs are required to be expanded to prove the optimality of solutions (task (ii)), and all NESs cannot possibly lead to optimal solutions and are thus irrelevant for achieving both task (i) and task (ii). Lemma 1. Given a problem instance I ICON, any admissible MOS Algorithm A must expand all MESs (Def. 1).

Proof. Let POCI, be the set of POF path costs of instance I. We need to show that every state s for which there exists a Pareto-optimal path U from start to s and there exists p POCI s.t. p c(U) + h(s), then s must be expanded by Algorithm A. We prove the contrapositive. Suppose that (i) I1 = G1 = V, E, c1 , start, goal, h is a problem instance in ICON whose POF has a set of costs POCI1, (ii) U is a Pareto-optimal path from start to s in I1 such that C c1(U) + h(s) for all C POCI1, (i.e., c1(U) + h(s) is in Zone A B), and (iii) A is an admissible DXBB algorithm that returns a POF for I1, denoted A(I1) (with cost vectors POCI1), without expanding s. Then, construct a new instance I2 IAD whose POF

has a set of cost vectors POCI2, such that (i) there exists C POCI2 for which C C for all C POCI1, and (ii) A also returns A(I1) (that contains no solution of cost C), thereby showing that A is not admissible on I2. I2 = G2 = V, E {e = s, goal }, c2 , start, goal . G2 is copied from G1, with the following changes: (i) an edge from s to goal is added (if it didn t exist before) and (ii) cost c2 is identical to c1 with a single modification, c2(e) = h(s). Edge e from s to goal generates a solution path U = [start, . . . , s, goal] whose total cost is c2(U ) = c1(U) + h(s), which is undominated by any cost in POCI1 (by assumption (ii)). Edge e is thus an essential part of any POF of I2. Since A is DXBB, it will behave on I2 similarly to its behavior on I1. In particular, A will not expand s, will not discover edge e, and will incorrectly return A(I1) as the POF for I2. Hence, if I2 IAD, A is not admissible. To complete the proof we now show that I2 IAD, i.e., that h is admissible on G2 (even after edge e was added). Let x be an arbitrary state in G2 and let W be any acyclic path in G2 from x to goal. We need to prove that h(x) c2(W). If W does not contain the new edge e, the claim trivially follows from the admissibility of h on G1. Now, suppose W does contain e. In other words, W = {X, e} for some path X from x to s (i.e., e is the last edge of W). Now, since X is a path from x to s in G1 and h is consistent on G1 we have: h(x) c1(X) + h(s) By definition, c2(e) = h(s) and c1(X) = c2(X), thus: = c2(X) + c2(e) = c2(W) Therefore, h is admissible on G2, i.e., I2 IAD.

We next show that NESs should never be expanded by admissible search algorithms. Lemma 2. Given an instance I IAD, every NES (Def. 2) cannot lead to any path with cost p POCI.

Proof. Assume by contradiction that (i) there exists a problem instance I such that the costs of all paths from start to some state s in I are dominated by one of the values in POCI and that (ii) s must be expanded by any admissible MOS Algorithm A in order to find a POF of I (denoted as A(I)). Let W = XY be an acyclic path from start to goal such that X is a path from start to s and Y is a path from s to goal. Due to the admissibility of I, we know that h(s) c(Y ) for all paths Y from s to goal. Thus, we get: c(W) = c(X) + c(Y ) c(X) + h(s) and, due to the assumption: p for some p POCI Thus, all paths W that pass through s are not a POF solution, and s should not be expanded.

4 Moving from States to Nodes Above, we classified states. Nonetheless, search algorithms typically operate in the context of nodes, which correspond to specific paths in the graph. In SOS with a consistent heuristic, node f-values are monotonically non-decreasing along paths. For states with multiple paths, A uses duplicate detection, associating each state s with a unique node n(s) representing

Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence (IJCAI-24)

Algorithm 1: MOS-A

Input : G = S, E, c , start, goal, h, OF, TB

1 nroot new node at start with g(nroot) = 0, h(nroot) = h(start), and parent(nroot) = null

2 initialize OPEN with OF and TB

3 OPEN.insert(nroot); POF

4 while OPEN = do

5 n OPEN.pop() // using OF and TB

6 if dom check exp(n) then continue

7 if s(n) = goal then

8 add n to POF and continue // line 4

9 for s S s.t. (s(n), s ) E do

10 n new node at s with h(n ) = h(s ), g(n ) = g(n) + c((s, s )), parent(n ) = n

11 if dom check gen(n ) then continue // line 9

12 OPEN.insert(n )

13 return POF

the cost-minimal path discovered for s, while pruning the remaining paths. Thus, choosing to expand a node n(s) with minimal f-value in OPEN ensures that n(s) represents an optimal path in the graph from start to s [Dechter and Pearl, 1985]. These properties allow the adaptation of the state classification to node classification in SOS. Given a heuristic function h, A (i) must expand all nodes n with f(n) < C , (ii) should never expand a node with n with f(n) > C and (iii) might expand some nodes with f(n) = C . In A all nodes in OPEN with the minimal f-value are said to be at the front of OPEN and A must choose for expansion a node from the front. Since there could be several nodes that have the same minimal f-value, a tie-breaking policy (TB) is applied to choose a specific node among them. To make a similar transition from states to nodes in MOS, we first generalize A to MOS-A (applicable for MOS).

4.1 Multi-Objective Search A (MOS-A )

In MOS-A , OPEN nodes are associated with f-values, which are d-dimensional vectors of values. Thus, in MOSA , the front of OPEN includes all nodes with undominated (but may be equal, relation) f-vectors. front can include multiple f-vectors (e.g., (1, 2) and (2, 1)), each can be used for choosing nodes for expansion. Thus, to choose a node from the front of OPEN, we have two phases. The first phase is unique to MOS (and does not exist in SOS) and is called an Ordering Function (OF) [Skyler et al., 2022]. An OF defines a total order between the different f-values. The cluster of nodes with the minimal order (according to the OF) is first chosen. A common OF is the lexicographic order, used by many algorithms, because it has a significant advantage in its CPU time overhead and in its data structure used [Hern andez et al., 2023b]. We discuss and compare this and other OFs below. The second phase uses TB within the chosen cluster (if it has more than one node) to choose the final node to expand. The pseudocode for the MOS-A framework is shown in Algorithm 1. It accepts a MOS instance as well as an OF and a TB. Each node n contains a state s(n), and a g-value vector g(n). f(n) is then calculated as f(n) = g(n) + h(s(n)). The front of OPEN includes all nodes whose f(n)-values are undominated (but may be equal, relation). In each itera-

tion, MOS-A extracts a node n with the best f-value from the front of OPEN, as determined by OF and TB (Line 5). MOS-A then performs the common expansion cycle and either adds a node to POF if s(n) = goal (Lines 7-8) or expands it (Lines 9-12). When expanding a node n, MOS-A generates a child node for each successor s of s(n). When OPEN becomes empty, it terminates and returns the POF.

4.2 Dominance Checks in MOS-A

MOS-A performs dominance checks for node n to determine whether n has the potential to be included in POF. Otherwise, n is pruned if it does not pass the check. Two types of checks need to be performed on node n. The first is called path dominance check. Here, we compare different occurrences of s(n) and prune those occurrences whose path costs are weakly dominated by costs of other occurrences. After this, only undominated nodes are left in OPEN and CLOSED. This is a generalization of duplicate detection in SOS, where only one node is kept for each state with minimal path cost. Similarly to SOS, this step is optional One may choose not to implement it and risk a large increase in the size of OPEN. This directly translates into CPU time and memory increase. The tradeoff is that path dominance check requires tailored data structures. The second type is called solution dominance check. Here, we prune node n if f(n) is dominated by a solution that was already found. This dominance check is mandatory. Without it, the algorithm can run forever and continue to produce solutions. This check is not performed in SOS because the search usually halts after finding the first solution. But, it is equivalent to pruning nodes with f > C , for example, if the task of the search is to return all optimal solutions. In MOS-A dominance checks can be performed in two locations: after a node was chosen for expansion (Line 6) or after a node was generated (Line 11) but before it is inserted into OPEN. Many algorithms use this framework (e.g., MOA [Stewart and White III, 1991], NAMOA [Mandow and De La Cruz, 2005], NAMOA -dr [Pulido et al., 2015], and BOA [Hern andez et al., 2023a]). While they all use lexicographic OF, they mainly differ in whether, when, and how dominance checks are implemented as well as what data structures are used for maintaining OPEN. These differences significantly impact the constant time per node expansion, leading to substantial speedup possibilities in CPU time. In our analysis, we focus solely on understanding the impact of various OFs and TBs on the number of node expansions, excluding considerations of CPU runtime. We now show how to adapt the classification of states to the classification of nodes for this family of MOS-A algorithms.

4.3 Node Classification in MOS-A

In MOS-A , any state s is represented by a set of nodes N(s), where each node n N(s) corresponds to an undominated path from start to s. Similarly to SOS, it was shown that given a consistent heuristic, the f-value of nodes is monotonically non-decreasing [Stewart and White III, 1991] along any given path. Thus, only expanding a node n N(s) for which f(n) is undominated by any other f-values in OPEN ensures

Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence (IJCAI-24)

Figure 2: Information propagation back from the goal.

that n represents a Pareto-optimal path from start to s. Consequently, the classification of states can be adapted to nodes, as we now show for MOS in general, but only illustrate for the special case of BOS in Figure 1.

Definition 4 (Must-Expand Nodes (MENs)). Given a problem instance I ICON, a node n is a MEN if f(n) is undominated by any cost in any POF of instance I, POCI, i.e., p POCI it holds that p f(n) .

Here too, the definition is split into two subcategories: Definition 4.1 (Optimality Nodes). Optimality nodes n dominate one solution (or more) from any POF, i.e., p POC such that f(n) p. Optimality nodes (Zone A in Figure 1) are analogous to the MENs in SOS (with f(n) < C ) and must be expanded to ensure the optimality of solutions in any POF. Definition 4.2 (Completeness Nodes). Completeness nodes n are mutually undominated by any solution in any POF, i.e., p POC it holds that f(n) p and p f(n). Completeness nodes (Zone B in Figure 1) have no analogy in SOS and have to be expanded to ensure there are no more solutions in any POF (i.e., that the POF is complete).

Definition 5 (Never-Expand Nodes (NENs)). NENs n are dominated by at least one path in any POF, i.e., p POC, such that p f(n).

NENs are analogous to nodes with f(n) > C in SOS and belong to Zone D which also includes the (dashed) border with Zone B in Figure 1. We note that the definition of MENs, albeit without the subdivision into two categories, and NENs is similar to the definitions presented by MPC10.

Definition 6 (Maybe-Expand Nodes (MBENs)). MBENs are nodes n that are neither MENs nor NENs, i.e, p POC such that f(n) = p.

All nodes in C points in Figure 1 are MBENs, which are analogous to nodes with f(n) = C in SOS. The proofs of Lemmas 1 (for MES) and 2 (for NES) can be directly adapted for nodes when using MOS-A and the above definitions.

Lemma 3. Any MOS-A algorithm expands all MENs and does not expand NENs.

Proof. Since MOS-A algorithms expand only nodes with undominated f-values and prune only dominated nodes, they are guaranteed to expand all MENs. We thus need to prove that MOS-A algorithms do not expand NENs. Assume by contradiction that a MOS-A Algorithm A expands a NEN n. Since n is a NEN, there exists a Pareto-optimal solution π such that π f(n). Since MOS-A algorithms expand only nodes with undominated f-values, π would be discovered before n is expanded. Thus, n will be pruned during dominance check, in contradiction to n being expanded by A.

Figure 3: Tie-breaking example in MOS-BF*.

4.4 State-Based vs. Node-Based Perspectives In the context of MOS, the definitions of Must-Expand (ME), Maybe-Expand (MBE), and Never-Expand (NE) are inherently different for nodes and states. With a consistent heuristic, the number of states expanded is bounded by the size of the graph. By contrast, the number of nodes representing cost-unique paths can be exponential in the number of states. Generally, the size of the Pareto-optimal frontier can be exponential in the number of states and thus counting states expanded may significantly underestimate the time and memory required by the algorithm. By contrast, considering node expansions can be too restrictive, potentially overestimating the resources needed by the search as we now demonstrate using Figure 2. There are k states, a1, . . . , ak, connected to s (start). Each of these k states is further linked to a series of states labeled b1, . . . , bn = g, forming k Pareto-optimal paths: s, ai, b1, . . . , nn 1, g, for 1 i k. In currently existing MOS-A algorithms, the search is carried out through distinct search nodes (paths) independently. Thus, each node ai (1 i k) will entail a distinct expansion of state b1, . . . , g (forming a distinct node for each of them). In this context, the definitions of MENs and NENs hold significance. However, under the DXBB assumption, an algorithm A may first expand s, a1, b1, . . . , g and discover the first Paretooptimal path. A can then propagate information, e.g., the path and its costs backward from g to b1. Since the expansion function is defined on states (and not paths), when a2 is expanded, the path from b1 to the goal is already known. As a result, b1, . . . , bn need not be expanded again. This also applies to the expansion of a3, . . . , ak. In these scenarios, where certain states need not be re-expanded, MENs and NENs may overestimate the number of expanded nodes where as MESs and NESs remain valid. We note that algorithms with such backpropagation and information sharing do not yet exist and are the focus of an active research effort. We thus next examine the influence of TBs and OFs in the context of MEN, MBEN, and NEN.

4.5 Role of Tie-Breaking Policies in MOS-A

Since MOS-A must expand all MENs and never expands any NEN, OFs and TBs can only affect which of the MBENs are expanded during the search. We discuss these next. TB policies determine which node to expand, among all nodes with the same f-value in OPEN. Consequently, different TBs can result in different nodes being expanded, even when the same OF is used. An example of the TBs effect on the search is shown in Figure 3. There are three states, start, x, and goal. Any MOS-A must expand start first (with any OF), generating nodes x and goal (both have the same f = (2, 2)). Thus, the order by which they are expended is solely determined by the TB. Consider the following two

Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence (IJCAI-24)

f1 max1 min1

(a) Normalization

2 3 4 5 . .. (b) Lex1

f1 (c) Lex2

1 2 3 4 5 . . .

1 2 3 4 5 . . .

4 1 2 3 4 5

f1 1 2 3 4 5 . .. (f) Max

Figure 4: The progress of different ordering functions. The three phases are colored differently.

TBs: ph which expand nodes in increasing order of their sum of h-values, and ph which expand nodes in decreasing order of their sum of h-values. Using ph , goal is expanded first (after start), finding a solution of cost (2, 2). In this case, the algorithm can terminate without expanding the node of x, as a solution that weakly dominates this node has been found (using the solution dominance check in Line 7 of Algorithm 1). By contrast, ph expands x and only then expands goal, resulting in three node expansions instead of two.

4.6 Role of Ordering Functions in MOS-A

Different TBs can result in different node expansions. We now show that using different ordering functions (OFs) with the same TB results in the same set of node expansions. Lemma 4. Let O1 and O2 be two OFs and let tb be a deterministic TB. A MOS-A Algorithm A that expands nodes according to O1 and tb will expand the same set of nodes on problem instance I ICON when using O2 and tb.

Proof. We prove the contrapositive. Let n be the first node expanded by A using O1 and not expanded using O2. Since A is admissible, then n is not a MEN. Otherwise, A must expand n even when using O2. In addition, n is not a NEN (Lemma 3). Thus, n is a MBEN and f(n) POC. A could avoid expanding n only by finding a solution of cost f(n) that does not pass through n. That solution must be discovered by O2 by expanding some other node n that is not expanded by O1, with f(n ) f(n) (due to admissibility). However, if f(n ) f(n), n would have been expanded before n when using O1 and n would have been pruned due to the solution dominance check, contradicting the fact that n was expanded using O1. Thus, f(n ) = f(n). Since n is the first node not expanded by A using O2, both n and n must have been in OPEN before either is expanded. The same deterministic tie breaking tb is used in both runs. Thus, the order by which n and n are expanded is identical in the two runs, in contradiction to n being expanded before n when using O1.

5 Ordering Functions While ordering functions (OFs) do not change the final set of nodes expanded, they can affect the order by which these nodes are expanded and thus the order and rate by which solutions are discovered. This may be crucial in anytime scenarios, where one needs many solutions as fast as possible. Frequently, the objectives are measured on different scales, e.g., time in seconds and distance in miles. Thus, to compare them meaningfully by OFs, we normalize them to be

on the same [0, 1] scale. Skyler et al. (2022) introduced a normalization scheme for BOS, which can also be used for MOS as we now describe. Let mini = minc POC ci and maxi = maxc POC ci. All the values in range [mini, maxi] are then linearly projected into range [0, 1]. This is achieved by the following projection function applied on any value xi: xi = xi mini maxi mini This normalization is illustrated for BOS in Figure 4a, where the gray and white circles are the solutions before and after the normalization, respectively. For BOS, mini and maxi values are calculated by running a SOS algorithm for objective i until all optimal solutions are found. mini is set to the cost of the optimal solution. maxj (for the other dimension i = j) is set to be the maximal value of objective j in all optimal solution costs for objective i. This procedure can only be done for BOS (with d = 2). In MOS with d > 2, this is not possible. For example, consider an instance where POC = {(2, 4, 10), (3, 3, 100), (4, 2, 10)}. Applying this method (i.e., running A for each objective) would only find {(2, 4, 10), (4, 2, 10)}. The maxi vector will be (4, 4, 10) instead of (4, 4, 100). So, in practice, one can take any upper (lower) bound on the maximal (minimal) solution for maxi (mini). Then, the normalized value x can be calculated based on these bounds using the normalization. We next describe several OFs and analyze their behavior. Figure 4 illustrates the progress of the different OFs on the search in the bi-objective search (BOS) setting. The search is divided into three phases: (1) The First phase nodes that are expanded before the first solution is found (the light grey area); (2) The Mid phase nodes that are expanded after finding the first solution but before finding the last solution (the dark grey area); and (3) The Last phase nodes that are expended after the last solution was found (the black area). The arrows indicate the order by which nodes are expanded.

Lexicographical Ordering (Lex). Given a MOS with d objectives, there are d! Lex orderings as follows. Let p = p1, . . . , pd be a permutation of [1, . . . , d]. Then, the total order of the f-values is based on the lexicographical order induced by p (i.e., first compare fp1, in the case of a tie, compare fp2, and so on). Figures 4b and 4c show Lex1 and Lex2 (resp.) for BOS. Lex1 finds the top-left solution first while Lex2 finds the bottom right solution first.

(Weighted) Average Ordering (Avg). Assume two nodes n and m with f-values f(n) = [fn1, fn2, . . . , fnd] and f(m) = [fm1, fm2, . . . , fmd], respectively. Let w be a weight vector. Avg is defined as follows:

Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence (IJCAI-24)

d = 2 d = 3 d = 4 Phase First Mid Last First Mid Last First Mid Last Lex1 17.2 82.8 0.0 12.7 87.3 0.0 9.9 90.1 0.0 Lex2 17.1 82.9 0.0 12.1 87.9 0.0 9.5 90.1 0.4 Avg 47.1 51.4 1.5 25.1 74.6 0.3 19.6 80.1 0.3 Min 27.9 68.6 3.5 14.6 82.3 3.1 11.0 86.5 2.5 Max 43.1 56.9 0.0 28.5 71.5 0.0 20.2 79.8 0.0

Table 1: Percentage of expansions in each search stage.

Avg(n, m) = arg min Pd i=1 wi fni , Pd i=1 wi fmi

If w = [1, . . . , 1], then it is considered to be the sum OF. A similar order would result from w = [1/d, . . . , 1/d], which is the mathematical average of the different fi-values. Figure 4d illustrates Avg. The f-values of expanded nodes form a diagonal line that proceeds along the direction of the black arrow. Assuming the f-values are normalized, the slope of this diagonal line is defined by w. Maximum (Minimum) Ordering (Max, Resp. Min). First, Max (Min) orders the (normalized) objectives of each node in decreasing (increasing) order. Then, the ordered objectives are compared lexicographically. Figures 4f and 4e present Max and Min OFs, respectively.

6 Empirical Evaluation We evaluated the expansions of each OF in the different expansion phases defined above on 200 random instances of the BAY road-map [DIMACS, 2006] with 2 4 objectives: time, distance, money, and uniform cost (respectively).

6.1 Expansions in the Three Phases Table 1 shows the ratio (in %) of expansions in each of the three phases compared to the entire run. For d = 2, in the First phase, as expected, Lex1 and Lex2 have a small percentage. This is because they find the (lexicographically) first solution with a perfect heuristic in the first dimension. Min alternates between Lex1 and Lex2, and finds the first solution after expanding slightly less than the sum of the two Lex orderings as it expands their intersection only once. Finally, Avg and Max expanded almost 50% of the nodes in the First phase. The Mid phase is the heaviest phase as it finds all solutions except the first one. In the Last phase, the nodes are expanded to prove that there are no more solutions in the POF. Min expanded 3.5% of the nodes after the last solution was found (black zone in Figure 4). This means that Min found the entire POF the fastest, as all functions expanded the same set of nodes in each instance. Importantly, Lex and Max do not expand any node in the Last phase, and the last solution found is the last node expanded for these OFs. The rest columns present these measurements for d = 3 and d = 4. The same trends are shown, but the relative size of the Mid phase becomes larger because there are more solutions to find.

6.2 Ratio of Solution Found Figure 5 presents another view of this experiment by showing the average ratio of solutions found from the POF (yaxis) as a function of the progress of the expansions (x-axis), measured in percentiles for BOS problems (d = 2). There is a curve for each of the OFs described above (Lex1, Lex2,

Percentage of Solutions Found

0 10 20 30 40 50 60 70 80 90 100 Percentage of Expansions

𝑀𝑀𝑀𝑀𝑀𝑀𝑚𝑚𝑚𝑚𝑚𝑚 𝐴𝐴𝐴𝐴𝐴𝐴𝑚𝑚𝑚𝑚𝑚𝑚 𝐿𝐿𝐿𝐿𝐿𝐿2 𝑚𝑚𝑚𝑚𝑚𝑚 𝐿𝐿𝐿𝐿𝐿𝐿1 𝑚𝑚𝑚𝑚𝑚𝑚 𝑀𝑀𝑎𝑎𝑎𝑎(𝑚𝑚𝑚𝑚𝑚𝑚)

𝑀𝑀𝑀𝑀𝑀𝑀 𝐴𝐴𝐴𝐴𝐴𝐴 𝐿𝐿𝐿𝐿𝐿𝐿2 𝐿𝐿𝐿𝐿𝐿𝐿1 𝑀𝑀𝑀𝑀𝑀𝑀

{ $ \ o le x i[ 1 ]

Figure 5: Solutions found vs. percentage of expansions.

Min, Avg, Max). Being close to the top left corner is desirable because this indicates that many solutions are found early during the search. The results show that on average, Min was earliest in finding larger portion of the POF, then Lex1 and Lex2 while Max and Avg were the slowest. This is reasonable as many completeness nodes (Zone B) are often located in the center of the POF (can also be seen in Figure 1), which is the last area Min explores. Additionally, the dots at the top of the figure present the maximal value (among the different runs) that the OF achieved for each percentile (Lexi(max), Min(max), Avg(max), Max(max)). The maximum values correlate with the average results. As mentioned, the last node explored by Lex and Max is a solution. Therefore, these functions reach 100% of the solutions in the end, after expanding all 100% of the nodes. By contrast, there was an instance in which Avg found all POF after 73% of the expansions, and Min found the entire POF after only 43% of the expansions. To summarize, while Lex reaches the first solution the fastest, other functions (Avg and Min) are able to find the entire POF faster, with fewer expansions. The same trends were seen for d = 3 and d = 4. These results are reported in the supplementary material. Note that the different ordering, together with the compatible implementation of dominance checks, may significantly reduce the CPU time per node expansion and the overall running time. Improvements of up to a few orders of magnitude were observed by the method that use Lex and the dimensionality reduction that is connected to it [Hern andez et al., 2023a]. These dimensionality reductions are also possible for Min, Max, and Avg, and are the matter of current ongoing but orthogonal work. Nevertheless, in this paper we solely focus the number of nodes expanded. We leave the discussion on actual CPU time which is a function of the constant time per node of the different implementations to a different paper.

7 Conclusions

We analyzed expansions in MOS, classifying them to Must Expand, Maybe-Expand, and Never-Expand, both in the context of states and nodes. In addition, we considered the issue of OFs, which are used by MOS-A algorithms to decide which node to expand next based on their f-values. We presented several OFs and compared them experimentally, showing the rate OFs find solutions. As all OFs must expand the same set of nodes, we showed the benefit of some OFs with regard to the order in which these nodes are explored.

Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence (IJCAI-24)

Acknowledgements

The research was supported by the United States-Israel Binational Science Foundation (BSF) under grant number 2021643 and by Israel Science Foundation (ISF) under grant number 909/23. The research at the University of Southern California was also supported by the National Science Foundation (NSF) under grant numbers 1817189, 1837779, 1935712, 2121028, 2112533, and 2321786. Finaly, the research was also supported by the National Center for Artificial Intelligence (CENIA), Financiamiento Basal ANID, and the Centro Ciencia & Vida under grant numbers FB210017, FB210008. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the sponsoring organizations, agencies, or any government.

[Ahmadi et al., 2021] Saman Ahmadi, Guido Tack, Daniel Harabor, and Philip Kilby. Bi-Objective Search with Bi Directional A . In Petra Mutzel, Rasmus Pagh, and Grzegorz Herman, editors, 29th Annual European Symposium on Algorithms, volume 204 of LIPIcs, pages 3:1 3:15, 2021.

[Bachmann et al., 2018] Daniel Bachmann, Fritz B okler, Jakob Kopec, Kira Popp, Bj orn Schwarze, and Frank Weichert. Multi-Objective Optimisation Based Planning of Power-Line Grid Expansions. ISPRS International Journal of Geo-Information, 7(7):258, 2018.

[Cl ımaco and Pascoal, 2012] J. Cl ımaco and M. Pascoal. Multicriteria Path and Tree Problems: Discussion on Exact Algorithms and Applications. ITOR, 19(1-2):63 98, 2012.

[Current and Marsh, 1993] J. Current and M. Marsh. Multiobjective Transportation Network Design and Routing Problems: Taxonomy and Annotation. EJOR, 65(1):4 19, 1993.

[Dechter and Pearl, 1985] Rina Dechter and Judea Pearl. Generalized Best-First Search Strategies and the Optimality of A . J. ACM, 32(3):505 536, 1985.

[DIMACS, 2006] DIMACS. The 9th DIMACS Implementation Challenge: Shortest Path, 2006.

[Eckerle et al., 2017] J urgen Eckerle, Jingwei Chen, Nathan R. Sturtevant, Sandra Zilles, and Robert C. Holte. Sufficient Conditions for Node Expansion in Bidirectional Heuristic Search. In ICAPS, pages 79 87, 2017.

[Fu et al., 2019] Mengyu Fu, Alan Kuntz, Oren Salzman, and Ron Alterovitz. Toward Asymptotically-Optimal Inspection Planning Via Efficient Near-Optimal Graph Search. In Robotics: Science and Systems, 2019.

[Fu et al., 2021] Mengyu Fu, Oren Salzman, and Ron Alterovitz. Computationally-Efficient Roadmap-based Inspection Planning via Incremental Lazy Search. In ICRA, pages 7449 7456, 2021.

[Goldin and Salzman, 2021] Boris Goldin and Oren Salzman. Approximate Bi-Criteria Search by Efficient Representation of Subsets of the Pareto-Optimal Frontier. In ICAPS, pages 149 158, 2021. [Hart et al., 1968] Peter E Hart, Nils J Nilsson, and Bertram Raphael. A Formal Basis for the Heuristic Determination of Minimum Cost Paths. IEEE Transactions on Systems Science and Cybernetics, 4(2):100 107, 1968. [Hern andez et al., 2023a] Carlos Hern andez, William Yeoh, Jorge Baier, Ariel Felner, Oren Salzman, Han Zhang, Shao-Hung Chan, and Sven Koenig. Multi-objective search via lazy and efficient dominance checks. In IJCAI, 2023. [Hern andez et al., 2023b] Carlos Hern andez, William Yeoh, Jorge A. Baier, Han Zhang, Luis Suazo, Sven Koenig, and Oren Salzman. Simple and efficient bi-objective search algorithms via fast dominance checks. Artif. Intell., 314:103807, 2023. [Mandow and De La Cruz, 2005] Lawrence Mandow and Jos e Luis P erez De La Cruz. A New Approach to Multi Objective A Search. In IJCAI, pages 218 223, 2005. [Mandow and P erez-de-la-Cruz, 2010] Lawrence Mandow and Jos e-Luis P erez-de-la-Cruz. Multiobjective A search with consistent heuristics. J. ACM, 57(5):27:1 27:25, 2010. [Pulido et al., 2015] Francisco-Javier Pulido, Lawrence Mandow, and Jos e-Luis P erez-de-la Cruz. Dimensionality Reduction in Multiobjective Shortest Path Search. CAR, 64:60 70, 2015. [Salzman et al., 2023] Oren Salzman, Ariel Felner, Carlos Hern andez, Han Zhang, Shao-Hung Chan, and Sven Koenig. Heuristic-search approaches for the multiobjective shortest-path problem: Progress and research opportunities. In IJCAI, pages 6759 6768, 2023. [Skriver, 2000] A. J.V. Skriver. A Classification of Bicriterion Shortest Path (BSP) Algorithms. Asia Pacific J. of Oper. Res., 17(2):199 212, 2000. [Skyler et al., 2022] S. Skyler, D. Atzmon, A. Felner, O. Salzman, H. Zhang, S. Koenig, W. Yeoh, and C. Hern andez. Bounded-Cost Bi-Objective Heuristic Search. In So CS22, pages 239 243, 2022. [Stewart and White III, 1991] Bradley S Stewart and Chelsea C White III. Multiobjective A . Journal of the ACM, 38(4):775 814, 1991. [Tarapata, 2007] Z. M. Tarapata. Selected Multicriteria Shortest Path Problems: An Analysis of Complexity, Models and Adaptation of Standard Algorithms. IJAM, 17(2), 2007. [Ulungu and Teghem, 1991] EL Ulungu and J Teghem. Multi-Objective Shortest Path Problem: A Survey. In Workshop on Multicriteria Decision Making, pages 176 188, 1991.

Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence (IJCAI-24)