Generalised Reachability Games: Complexity & Strategies

• Generalised reachability games are turn-based graph games where players meet conjunctive objectives by visiting a family of target sets in any order.
• They exhibit a complexity spectrum from fixed-parameter tractability to PSPACE-completeness, depending on the structure and size of the target sets.
• Algorithmic strategies leverage product constructions, attractor methods, and precise memory bounds to guide effective strategy synthesis.

Generalised reachability games are turn-based graph games in which the winning condition is not a single target set but a family of target sets that must all be visited during the play, in arbitrary order. In the standard finite two-player setting, the arena is a directed graph whose vertices are partitioned between Eve and Adam; Eve seeks to ensure that every target set $F_i$ is hit at least once, while Adam seeks to prevent this. This objective is the conjunction of reachability conditions, and it has become a canonical benchmark for the boundary between tractable and intractable multi-goal synthesis: the unrestricted problem is $\PSPACE$-complete, yet several structurally important fragments are polynomial-time solvable or fixed-parameter tractable (Fijalkow et al., 2010, Bose et al., 17 Sep 2025). Subsequent work has refined this picture for stochastic, temporal, hybrid, symbolic, and optimisation variants, and has clarified which extensions preserve decidability, approximation, or effective strategy synthesis (Ashok et al., 2019, Austin et al., 2024, Chatterjee et al., 11 May 2026).

1. Formal objective and semantic variants

In the classical setting, games are played on an arena $A = (V, (V_\text{Eve}, V_\text{Adam}), E)$, and a play is an infinite path $v_0,v_1,\ldots$ with $(v_i,v_{i+1}) \in E$. For a single target set $F \subseteq V$, the reachability objective is

$$\mathrm{Reach}(F)=\{v_0,v_1,v_2,\ldots \mid \exists p \in \mathbb{N},\, v_p \in F\}.$$

Given $k$ target sets $F_1,\ldots,F_k$, the generalised reachability objective is

$$\mathrm{GenReach}(F_1,\ldots,F_k)=\left\{\pi \mid \forall i \in \{1,\ldots,k\},\, \exists p_i \in \mathbb{N}: v_{p_i} \in F_i\right\},$$

so Eve must visit at least one vertex from every $\PSPACE$0 (Fijalkow et al., 2010).

This formulation admits several immediately relevant specialisations. Standard reachability is the case $\PSPACE$1. Explorability is a singleton-target special case in which each objective is $\PSPACE$2, so the requirement is to visit every vertex; on temporal graphs, this special case turns out to have the same complexity as generalised reachability (Austin et al., 2024). In stochastic games, the qualitative objective is replaced by a vector of threshold guarantees: for target family $\PSPACE$3, a strategy achieves value vector $\PSPACE$4 if

$\PSPACE$5

and the central object becomes the achievable set and its Pareto frontier rather than a Boolean winning region (Ashok et al., 2019).

A persistent technical point is that generalised reachability is conjunctive, not sequential. No order between the $\PSPACE$6 is prescribed, and the target sets may overlap. This makes the objective more expressive than a single reachability condition, but it also means that the state of the game must encode which obligations have already been discharged. That observation underlies both the standard fixed-parameter algorithms and the memory bounds for winning strategies (Fijalkow et al., 2010).

2. Complexity landscape for finite turn-based games

The foundational complexity result is that solving generalised reachability games is $\PSPACE$7-complete (Fijalkow et al., 2010). The hardness proof is by reduction from QBF: existential and universal quantifiers are simulated by Eve and Adam choices, and clauses are encoded as reachability sets. Membership in $\PSPACE$8 follows from an alternating simulation argument together with the fact that if Eve can win, she can do so within $\PSPACE$9 steps, where $A = (V, (V_\text{Eve}, V_\text{Adam}), E)$0 is the number of vertices and $A = (V, (V_\text{Eve}, V_\text{Adam}), E)$1 the number of objectives (Fijalkow et al., 2010).

At the same time, the problem is fixed-parameter tractable in the number $A = (V, (V_\text{Eve}, V_\text{Adam}), E)$2 of objectives. The key observation is that the only relevant auxiliary information is which target sets have already been visited. A product construction with a $A = (V, (V_\text{Eve}, V_\text{Adam}), E)$3-state memory structure reduces the game to an ordinary reachability game, yielding running time

$A = (V, (V_\text{Eve}, V_\text{Adam}), E)$4

where $A = (V, (V_\text{Eve}, V_\text{Adam}), E)$5 is the number of edges (Fijalkow et al., 2010).

Later work sharpened the tractable frontier. If a game has $A = (V, (V_\text{Eve}, V_\text{Adam}), E)$6 singleton targets and only $A = (V, (V_\text{Eve}, V_\text{Adam}), E)$7 non-singleton targets, then GenReach is solvable in $A = (V, (V_\text{Eve}, V_\text{Adam}), E)$8, so it is polynomial-time whenever the number of non-singleton targets is logarithmic in the game size. The same work also identifies a polynomial-time case with exactly one large target set, extending the older singleton-only tractability boundary (Bose et al., 17 Sep 2025).

Fragment	Complexity status	Source
General two-player GenReach	$A = (V, (V_\text{Eve}, V_\text{Adam}), E)$9-complete	(Fijalkow et al., 2010)
Parameterised by number of objectives $v_0,v_1,\ldots$0	FPT, $v_0,v_1,\ldots$1	(Fijalkow et al., 2010)
$v_0,v_1,\ldots$2 singleton targets and $v_0,v_1,\ldots$3 non-singleton targets	$v_0,v_1,\ldots$4	(Bose et al., 17 Sep 2025)
All singleton targets	polynomial-time; reported as $v_0,v_1,\ldots$5-complete in later summary	(Bose et al., 17 Sep 2025)
Two-player, all target sets of size at most $v_0,v_1,\ldots$6	open	(Fijalkow et al., 2010, Bose et al., 17 Sep 2025)

The dependence on target-set size is unusually sharp. Earlier results showed that singleton target sets yield a polynomial-time fragment, while in the one-player setting size-2 target sets are still polynomial-time solvable, by reduction to 2-CNF satisfiability. In contrast, the exact complexity of the two-player size-2 case remains open, whereas target sets of size at least $v_0,v_1,\ldots$7 already recover the full $v_0,v_1,\ldots$8-hardness of the unrestricted problem (Fijalkow et al., 2010, Bose et al., 17 Sep 2025).

This makes generalised reachability a standard example of a problem that becomes hard under very mild-looking extensions of the specification language. The product reduction to ordinary reachability is algorithmically useful, but it does not eliminate the combinatorial blow-up associated with tracking multiple obligations.

3. Strategy structure, memory, and attractor methods

The same memory structure that yields the fixed-parameter algorithm also gives sharp strategy bounds. Eve can always win with at most $v_0,v_1,\ldots$9 memory states, and Adam can always win with at most $(v_i,v_{i+1}) \in E$0 memory states; both bounds are tight (Fijalkow et al., 2010). The lower-bound examples show that this is not merely an artefact of the product construction. In particular, Eve may need to distinguish all nonempty subsets of already-visited targets, while Adam may need to remember which obligation remains vulnerable (Fijalkow et al., 2010).

Algorithmically, the standard reduction is to take the product of the arena with a bitvector over $(v_i,v_{i+1}) \in E$1, recording which target sets have been met. Once all bits are set, the objective becomes ordinary reachability in the product arena. This explains both the exponential dependence on $(v_i,v_{i+1}) \in E$2 and the fact that memoryless winning strategies in the product correspond to finite-memory strategies in the original game (Fijalkow et al., 2010).

For structurally simpler fragments, attractor methods yield more specialised algorithms. Given a set $(v_i,v_{i+1}) \in E$3, the attractor is defined iteratively by

$(v_i,v_{i+1}) \in E$4

and $(v_i,v_{i+1}) \in E$5 (Bose et al., 17 Sep 2025). When all targets are singletons, inclusion relations between singleton attractors induce a preorder that supports polynomial-time algorithms. In the optimisation setting, the SCC DAG of this preorder can be combined with dynamic programming to compute best achievable singleton subsets (Bose et al., 17 Sep 2025).

A common misconception is that generalised reachability is “just reachability with bookkeeping.” The bookkeeping is precisely what drives both the upper bounds and the lower bounds: it is enough to obtain FPT algorithms, but it is also sufficient to force exponential memory for Eve and to separate tractable and intractable fragments.

4. Optimisation, stochasticity, and quantitative generalisations

Recent work has studied optimisation versions in which not all target sets need be guaranteed. In MaxGenReach, Eve asks whether she can force visiting at least $(v_i,v_{i+1}) \in E$6 target sets; in MaxGenReachPromise, she must first pledge a $(v_i,v_{i+1}) \in E$7-subset and then force all pledged targets. These variants behave differently from the classical decision problem. MaxGenReach is $(v_i,v_{i+1}) \in E$8-complete in the general two-player case, even when all target sets have size $(v_i,v_{i+1}) \in E$9, and it is already NP-hard when all target sets are singletons. By contrast, MaxGenReachPromise is polynomial-time solvable for singleton targets, but $F \subseteq V$0-complete in general (Bose et al., 17 Sep 2025). The distinction shows that “best-effort” synthesis is substantially more delicate than all-target synthesis.

For turn-based stochastic games with generalised-reachability objectives of arbitrary dimension, exact decidability of the threshold problem remains open, but the achievable value vectors can be approximated to any desired precision $F \subseteq V$1 (Ashok et al., 2019). The central result is an anytime algorithm that computes converging lower and upper polyhedral bounds $F \subseteq V$2 on the achievable set, with explicit directional error bounds. The update scheme generalises Bellman iteration from scalar values to convex downward-closed sets, and convergence of the upper bound requires regional deflation of end components, because the minimiser’s optimal exit can depend on the direction in which the Pareto frontier is probed (Ashok et al., 2019).

Quantitative reachability games form a related but distinct branch of the theory. Min-cost reachability games attach a cumulative weight to the prefix ending at the first visit of a target set $F \subseteq V$3: Min tries to reach $F \subseteq V$4 with minimal total cost, and Max tries to maximise this cost or prevent target visitation. These games admit a pseudo-polynomial value-iteration algorithm and serve as a core subroutine for total-payoff games (Brihaye et al., 2014). At a more abstract level, Galois energy games provide a generic framework for energy-bounded attacker games with reachability objective over well-founded bounded join-semilattices; the framework subsumes multi-weighted reachability games and yields algorithms polynomial in graph size and exponential in dimension for the cited instantiations (Lemke et al., 22 Apr 2025). This suggests a broader quantitative landscape in which conjunction of targets is only one way of strengthening classical reachability.

5. Infinite-state, hybrid, temporal, and symbolic settings

Generalised reachability acquires additional complexity when the underlying system is infinite-state or time-dependent. In o-minimal hybrid reachability games, two observation regimes are distinguished. Under perfect observation, time-abstract bisimulation is not adequate, and the correct abstraction is suffix equivalence based on trajectory encodings. Under partial observation, the corresponding abstraction is the superword encoding, which records, for each delay, the set of partition pieces that may be reached. For o-minimal hybrid systems with a decidable theory, these abstractions give finitely many equivalence classes, decidability, computability of winning sets, and computable memoryless strategies (0911.4833).

In infinite-state polynomial reachability games over real variables, a recent automated method introduces ranking certificates as a sound and complete proof rule for the single-target reachability objective, under finite branching for the SAFE player. A ranking certificate is a function $F \subseteq V$5 satisfying monotone decrease conditions tailored to SAFE and REACH states; once such an $F \subseteq V$6 exists, a memoryless winning strategy is obtained by choosing a ranking successor. The associated template-based synthesis procedure fixes polynomial templates for the certificate and the REACH strategy, generates quantified polynomial constraints, applies quantifier elimination, and solves the resulting existential system by SMT. The method is sound, semi-complete, and sub-exponential for fixed degree. However, the same work explicitly identifies extension to generalised reachability, including conjunctions of multiple reachability goals, as future research: the certificate method is specific to a single target set, and generalised reachability would require new multidimensional or lexicographic certificates (Chatterjee et al., 11 May 2026).

Temporal graphs expose another boundary. Explorability is the special case of generalised reachability in which each target set is a singleton vertex, yet on temporal graphs its complexity jumps to that of generalised reachability itself. For explicitly represented temporal graphs, one-player explorability and generalised reachability are $F \subseteq V$7-complete, while the two-player versions are $F \subseteq V$8-complete. For symbolically represented temporal graphs, even one-player reachability becomes $F \subseteq V$9-complete, and two-player explorability and generalised reachability are $$\mathrm{Reach}(F)=\{v_0,v_1,v_2,\ldots \mid \exists p \in \mathbb{N},\, v_p \in F\}.$$0-hard and in $$\mathrm{Reach}(F)=\{v_0,v_1,v_2,\ldots \mid \exists p \in \mathbb{N},\, v_p \in F\}.$$1 (Austin et al., 2024). The temporal dimension effectively converts visiting many vertices into a synchronisation problem over many time-indexed obligations.

These developments show that the finite-graph theory does not transfer mechanically. In some settings, such as o-minimal hybrid systems, trajectory-sensitive abstractions recover decidability; in others, such as polynomial games over reals, even the single-target case requires new certificate technology, and the generalised objective remains open.

6. Structured families, imperfect information, and open problems

The theory also interacts with symbolic variability and richer information structures. In software product lines, featured games annotate transitions with feature guards and support family-based analysis over all products simultaneously. Featured reachability games admit symbolic fixed-point algorithms based on guard partitions and late splitting, and the framework extends to minimum reachability, discounted, energy, and parity games while preserving optimal featured strategies that project to optimal per-product strategies (Fahrenberg et al., 2020). Although this is not a direct theory of GenReach, it is a natural variability-aware generalisation of game-based reachability analysis.

Epistemic and concurrent models change the strategy semantics rather than the target condition. Reachability games in Dynamic Epistemic Logic use action models to describe informational effects; controller synthesis is $$\mathrm{Reach}(F)=\{v_0,v_1,v_2,\ldots \mid \exists p \in \mathbb{N},\, v_p \in F\}.$$2-complete for public announcements, $$\mathrm{Reach}(F)=\{v_0,v_1,v_2,\ldots \mid \exists p \in \mathbb{N},\, v_p \in F\}.$$3-complete for public actions, decidable for propositional actions with bounded modal depth, and undecidable in full generality. In the distributed setting, propositional actions with two existential agents already lead to undecidability, while hierarchical information restores decidability (Maubert et al., 2019). For concurrent graph games more broadly, determinacy can be lifted from turn-based to concurrent settings when every local interaction is a determined game form; then Borel objectives, including reachability and parity, inherit deterministic winning or value-optimal strategies (Bordais et al., 2021).

Several open directions recur across the literature. The most explicit finite-state open problem is the exact complexity of two-player GenReach when every target set has size at most $$\mathrm{Reach}(F)=\{v_0,v_1,v_2,\ldots \mid \exists p \in \mathbb{N},\, v_p \in F\}.$$4 (Bose et al., 17 Sep 2025). In stochastic games, exact threshold decidability for generalised reachability remains unsettled despite the existence of arbitrarily precise Pareto-frontier approximation (Ashok et al., 2019). In infinite-state synthesis, ranking certificates presently cover only standard reachability, and their extension to conjunctions of reachability objectives, as well as to $$\mathrm{Reach}(F)=\{v_0,v_1,v_2,\ldots \mid \exists p \in \mathbb{N},\, v_p \in F\}.$$5-regular and quantitative objectives, is still future work (Chatterjee et al., 11 May 2026).

Taken together, these results place generalised reachability games at a central junction of automata theory, game theory, synthesis, and verification. The objective is simple to state, but it is rich enough to expose sharp transitions between polynomial time, fixed-parameter tractability, $$\mathrm{Reach}(F)=\{v_0,v_1,v_2,\ldots \mid \exists p \in \mathbb{N},\, v_p \in F\}.$$6- and $$\mathrm{Reach}(F)=\{v_0,v_1,v_2,\ldots \mid \exists p \in \mathbb{N},\, v_p \in F\}.$$7-completeness, approximation without exact decidability, and decidability that depends delicately on observation, target-set structure, and the ambient state space.