Imperfect-Information Games and Generalized Planning

1 Introduction

Automated planning is a fundamental problem in Artificial Intelligence. Given a deterministic dynamic system with a single known initial state and a goal condition, automated planning consists of finding a sequences of actions (the plan) to be performed by agents in order to achieve the goal [M. Ghallab and Traverso, 2008]. The application of this notion in real-dynamic worlds is limited, in many situations, by three facts: i) the number of objects is neither small nor predetermined, ii) the agent is limited by its observations, iii) the agent wants to realize a goal that extends over time. For example, a preprogrammed driverless car cannot know in advance the number of obstacles it will enter in a road, or the positions of the other cars not in its view, though it wants to realize, among other goals, that every time it sees an obstacle it avoids it. This has inspired research in recent years on generalized forms of planning including conditional planning in partially observable domains [Levesque, 1996; Rintanen, 2004], planning with incomplete information for temporally extended goals [De Giacomo and Vardi, 1999; Bertoli and Pistore, 2004] and generalized planning for multiple domains or infinite domains [Levesque, 2005; Srivastava et al., 2008; Bonet et al., 2009; Hu and Levesque, 2010; Hu and De Giacomo, 2011; Srivastava et al., 2011; Felli et al., 2012; Srivastava et al., 2015].

We use the following running example, taken from [Hu and Levesque, 2010], to illustrate a generalized form of planning:

Aniello Murano, Sasha Rubin, Antonio Di Stasio

Universit`a degli Studi di Napoli "Federico II" Naples, Italy first.last@unina.it

Example 1 (Tree Chopping). The goal is to chop down a tree, and store the axe. The number of chops needed to fell the tree is unknown, but a look-action checks whether the tree is up or down. Intuitively, a plan solving this problem alternates looking and chopping until the tree is seen to be down, and then stores the axe. This scenario can be formalized as a partially-observable planning problem on a single infinite domain (Example 2, Figure 1), or on the disjoint union of infinitely many finite domains (Section 3).

The standard approach to solve planning under partial observability for finite domains is to reduce them to planning under complete observability. This is done by using the belief-state construction that removes partial observability and passes to the belief-space [Goldman and Boddy, 1996; Bertoli et al., 2006]. The motivating problem of this work is to solve planning problems on infinite domains, and thus we are naturally lead to the problem of removing partialobservability from infinite domains.

In this paper we adopt the Formal Methods point of view and consider generalized planning as a game G of imperfect information, i.e., where the player under control has partial observability. The game G may be infinite, i.e., have infinitely many states.

Our technical contribution (Theorem 4.5) is a general sound and complete mathematical technique for removing imperfect information from a possibly infinite game G to get a game Gβ , possibly infinite, of perfect information. Our method builds on the classic belief-state construction [Reif, 1984; Goldman and Boddy, 1996; Raskin et al., 2007], also adopted in POMDPs [Kaelbling et al., 1998; LaValle, 2006]. 1 The classic belief-state construction fails for certain infinite games, see Example 3. We introduce a new component to the classic belief-state construction that isolates only those plays in the belief-space that correspond to plays in G. This new component is necessary and sufficient to solve the general case and capture all infinite games G.

We apply our technique to the decision problem that asks, given a game of imperfect information, if the player under control has a winning strategy (this corresponds to deciding if there is a plan for a given planning instance). We remark that we consider strategies and plans that may depend on the history of the observations, not just the last observation. Besides showing how to solve the running Tree Chopping example, we report two cases. The first case is planning under partial observability for temporally extended goals expressed in LTL in finite domains (or a finite set of infinite domains sharing the same observations). This case generalizes well-known results in the AI literature [De Giacomo and Vardi, 1999; Rintanen, 2004; Bertoli et al., 2006; Hu and De Giacomo, 2011; Felli et al., 2012]. The second case involves infinite domains. Note that because game solving is undecidable for computable infinite games (simply code the configuration space of a Turing Machine), solving games with infinite domains requires further computability assumptions. We focus on games generated by pushdown automata; these are infinite games that recently attracted the interest of the AI community [Murano and Perelli, 2015; Chen et al., 2016]. In particular these games have been solved assuming perfect information. By applying our technique, we extend their results to deal with imperfect information under the assumption that the stack remains observable (it is known that making the stack unobservable leads to undecidability [Azhar et al., 2001]).

1 However, our work considers nondeterminism rather than probability, and qualitative objectives rather than quantitative objectives.

2 Generalized-Planning Games

In this section we define generalized-planning (GP) games, known as games of imperfect information in the Formal Methods literature [Raskin et al., 2007], that capture many generalized forms of planning.

Informally, two players (agent and environment) play on a transition-system. Play proceeds in rounds. In each round, from the current state s of the transition-system, the agent observes obs(s) (some information about the current state), and picks an action a from the set of actions Ac, and then the environment picks an element of tr(s↪ a) (tr is the transition function of the transition-system) to become the new current state. Note that the players are asymmetric, i.e., the agent picks actions and the environment resolves non-determinism.

Notation. Write X! for the set of infinite sequences whose elements are from the set X, write X⇤ for the finite sequences, and X+ for the finite non-empty sequences. If ⇡ is a finite sequence then Last(⇡) denotes its last element. The positive integers are denoted N, and N0 := N [ ¶0♦.

Linear-temporal logic. We define LTL over a finite set of letters ⌃. 2 The formulas of LTL (over ⌃) are generated by the following grammar: ' ::= x ♣ ' ^ ' ♣ ≻' ♣ X' ♣ 'U' where x 2 ⌃. We introduce the usual abbreviations for, e.g., _↪ F. Formulas of LTL (over ⌃) are interpreted over infinite words ↵ 2 ⌃! . Define the satisfaction relation ♣= as follows: i) (↵↪ n) ♣= x iff ↵n = x; ii) (↵↪ n) ♣= '1 ^ '2 iff (↵↪ n) ♣= 'i for i = 1↪ 2; iii) (↵↪ n) ♣= ≻' iff it is not the case that (↵↪ n) ♣= '; iv) (↵↪ n) ♣= X' iff (↵↪ n+1) ♣= '; v) (↵↪ n) ♣= '1 U'2 iff there exists i ≥ n such that (↵↪ i) ♣= '2 and for all j 2 [n↪ i), (↵↪ j) ♣= '1. Write ↵ ♣= ' if (↵↪ 0) ♣= ' and say that ↵ satisfies the LTL formula '.

2 This is without loss of generality, since if LTL were defined over a set of atomic propositions AP we let ⌃ = 2AP and replace atoms p 2 AP by Wp2x x to get equivalent LTL formulas over ⌃.

Arenas. An arena of imperfect information, or simply an arena, is a tuple A = (S↪ I↪ Ac↪ tr↪Obs↪ obs), where S is a (possibly infinite) set of states, I ✓ S is the set of initial states, Ac is a finite set of actions, and tr : S⇥Ac ! 2S ∩¶;♦ is the transition function, Obs is a (possibly infinite) set of observations, and obs : S ! Obs, the observation function, maps each state to an observation. We extend tr to sets of states: for ; 6= Q ✓ S, let tr(Q↪ a) denote the set [q2Qtr(q↪ a).

Sets of the form obs-1 (x) for x 2 Obs are called observation sets. The set of all observation sets is denoted ObsSet. Non-empty subsets of observation sets are called belief-states. Informally, a belief-state is a subset of the states of the game that the play could be in after a given finite sequence of observations and actions.

Finite and finitely-branching. An arena is finite if S is finite, and infinite otherwise. An arena is finitely-branching if i) I is finite, and ii) for every s↪ a the cardinality of tr(s↪ a) is finite. Clearly, being finite implies being finitely-branching.

Strategies. A play in A is an infinite sequence ⇡ = s0a0s1a1s2a2 ▷ ▷ ▷ such that s0 2 I and for all i 2 N0, si+1 2 tr(si↪ ai). A history h = s0a0 ▷ ▷ ▷ sn-1an-1sn is a finite prefix of a play ending in a state. The set of plays is denoted Ply(A), and the set of histories is denoted Hist(A) (we drop A when it is clear from the context). For a history or play ⇡ = s0a0s1a1 ▷ ▷ ▷ write obs(⇡) for the sequence obs(s0)a0obs(s1)a1 ▷ ▷ ▷. A strategy (for the agent) is a function σ : Hist(A) ! Ac. A strategy is observational if obs(h) = obs(h0 ) implies σ(h) = σ(h0 ). In Section 3 we will briefly mention an alternative (but essentially equivalent) definition of observational strategy, i.e., as a function Obs+ ! Ac. We do not define strategies for the environment. A play ⇡ = s0a0s1a1 ▷ ▷ ▷ is consistent with a strategy σ if for all i 2 N we have that if h 2 Hist(A) is a prefix of ⇡, say h = s0a0 ▷ ▷ ▷ sn-1an-1sn, then σ(h) = an+1.

GP Games. A generalized-planning (GP) game, is a tuple G = hA↪Wi where the winning objective W ✓ Obs! is a set of infinite sequences of observation sets. A GP game with restriction is a tuple G = hA↪ W↪ F i where, in addition, F ✓ S! is the restriction. Note that unlike the winning objective, the restriction need not be closed under observations. A GP game is finite (resp. finitely branching) if the arena A is finite (resp. finitely branching).

Winning. A strategy σ is winning in G = hA↪Wi if for every play ⇡ 2 Ply(A) consistent with σ, we have that obs(⇡) 2 W. Similarly, a strategy is winning in G = hA↪W↪ Fi if for every play ⇡ 2 Ply(A) consistent with σ, if ⇡ 2 F then obs(⇡) 2 W. Note that a strategy is winning in hA↪W↪ Ply(A)i if and only if it is winning in hA↪Wi.

Solving a GP game. A central decision problem is the following, called solving a GP game: given a (finite representation of a) GP game of imperfect information G, decide if the agent has a winning observational-strategy.

GP games of perfect information. An arena/GP game has perfect information if Obs = S and obs(s) = s for all s. We thus suppress mentioning Obs and obs completely, e.g., we write A = (S↪ I↪ Ac↪ tr) and W↪ F ✓ S! . Note that in a GP game of perfect information every strategy is observational.

Example 2 (continued). We formalize the tree-chopping planning problem. Define the GP game Gchop = hAchop↪Wi where Achop = hS↪ I↪ Ac↪ tr↪ Obs↪ obsi, and:

• S = ¶down↪ success↪ failure♦[(¶uk♦⇥N0)[(¶up♦⇥ N),
• Ac = ¶chop↪ look↪ store♦, I = ¶uk♦ ⇥ N,
• tr is illustrated in Figure 1,
• Obs = ¶DN↪ X↪ ⇥↪ UK↪ UP♦,
• obs maps down 7! DN, (up↪ i) 7! UP for i 2 N, (uk↪ i) 7! UK for i 2 N0, failure 7! ⇥, and success 7!X, and
• the objective W is defined as ↵ 2 W iff ↵ ♣= FX.

The mentioned plan is formalized as the observationalstrategy σchop that maps any history ending in (uk↪ i) to look (for i 2 N0), (up↪ i) to chop (for i 2 N), down to store, and all others arbitrarily (say to store).

Note: σchop is a winning strategy, i.e., no matter which initial state the environment chooses, the strategy ensures that the play (it is unique because the rest of the GP game is deterministic) reaches the state success having observation X.

3 Generalized-Planning Games and Generalized Forms of Planning

In this section we establish that generalized-planning (GP) games can model many different types of planning from the AI literature, including a variety of generalized forms of planning:

1. planning on finite transition-systems, deterministic actions, actions with conditional effects, partially observable states, incomplete information on the initial state, and temporally extended goals [De Giacomo and Vardi, 1999];
2. planning under partial observability with finitely many state variables, nondeterministic actions, reachability goals, and partial observability [Rintanen, 2004];
3. planning on finite transition systems, nondeterministic actions, looking for strong plans (i.e., adversarial nondeterminism) [Bertoli et al., 2006];
4. generalized planning, consisting of multiple (possibly infinitely many) related finite planning instances [Hu and Levesque, 2010; Hu and De Giacomo, 2011].

We discuss the latter in detail. Following [Hu and De Giacomo, 2011], a generalized-planning problem P is defined as a sequence of related classical planning problems. In our terminology, fix finite sets Ac↪Obs and let P be a countable sequence G1↪G2↪ ▷ ▷ ▷ where each Gn is a finite GP game of the form hSn↪ ¶◆n♦↪ Ac↪ trn↪Obs↪ obsn↪Wni. In [Hu and De Giacomo, 2011], a plan is an observationalstrategy p : Obs+ ! Ac, and a solution is a single plan that solves all of the GP games in the sequence. Now, we view P as a single infinite GP game as follows. Let GP denote the disjoint union of the GP games in P. Formally, GP = hS↪ I↪ Ac↪ tr↪ Obs↪ obs↪ Wi where

• S = ¶(s↪ n) : s 2 Sn↪ n 2 N♦,
• I = ¶(◆n↪ n) : n 2 N♦,
• tr((s↪ n)↪ a) = ¶(t↪ n) : t 2 trn(s↪ a)♦,
• obs(s↪ n) = obsn(s),
• W = [nWn.

Then: there is a correspondence between solutions for P and winning observational-strategies in GP.

For example, consider the tree-chopping problem as formalized in [Hu and Levesque, 2010; Hu and De Giacomo, 2011]: there are infinitely many planning instances which are identical except for an integer parameter denoting the number of chops required to fell the tree. The objective for all instances is to fell the tree. Using the translation above we get a GP game with an infinite arena which resembles (and, in fact, can be transformed to) the GP game in Example 2.

4 Generalized Belief-State Construction

In this section we show how to remove imperfect information from generalized-planning (GP) games G. That is, we give a transformation of GP games of imperfect information G to GP games of perfect information Gβ such that the agent has a winning observational-strategy in G if and only if the agent has a winning strategy in Gβ . The translation is based on the classic belief-state construction [Reif, 1984; Raskin et al., 2007]. Thus, we begin with a recap of that construction.

Belief-state Arena. 3 Let A = (S↪ I↪ Ac↪ tr↪ Obs↪ obs) be an arena (not necessarily finite). Recall from Section 2 that observation sets are of the form obs-1 (x) for x 2 Obs, and are collectively denoted ObsSet. Define the arena of perfect information Aβ = (S β↪ I β↪ Ac↪ tr β) where,

• S β is the set of belief-states, i.e., the non-empty subsets of the observation-sets,
• I β consists of all belief-states of the form I \ X for X 2 ObsSet,
• tr β(Q↪ a) consists of all belief-states of the form tr(Q↪ a) \ X for X 2 ObsSet.

The idea is that Q 2 S β represents a refinement of the observation set: the agent, knowing the structure of G ahead of time, and the sequence of observations so far in the game, may deduce that it is in fact in a state from Q which may be a strict subset of its corresponding observation set X. 4

NB. Since Aβ is an arena, we can talk of its histories and plays. Although we defined S β to be the set of all beliefstates, only those belief-states that are reachable from I β are relevant. Thus, overload notation and write S β for the set of reachable belief-states, and Aβ for the corresponding arena. This notation has practical relevance since if A is countable there are uncountably many belief-states; but in many cases only countably many (or, as in the running example, finitely many) reachable belief-states.

The intuition for the rest of this section is illustrated in the next example.

Example 3 (continued). Figure 2 shows the arena Aβ chop corresponding to the arena from tree-chopping game Gchop, i.e.,

3 In the AI literature, this is sometimes called the belief-space.

4 To illustrate simply, suppose there is a unique initial state s, and that it is observationally equivalent to other states. At the beginning of the game the agent can deduce that it must be in s. Thus, its initial belief-state is ¶s♦ and not its observation-set obs-1 (obs(s)). This belief can (and, in general, must) be exploited if the agent is to win.

• S β are the following belief-states: ¶(uk↪ n) ♣ n 2 N0♦, denoted UK; ¶(up↪ n) ♣ n 2 N♦, denoted UP; ¶down♦, denoted DN; ¶success♦, denoted X; and ¶failure♦.
• I β is the belief-state UK,
• and tr β is shown in the figure.

Note that the agent does not have a winning strategy in the GP game with arena Aβ chop and winning condition FX. The informal reason is that the strategy σchop (which codifies "alternately look and chop until the tree is sensed to be down, and then store the axe"), which wins in G, does not work. The reason is that after every look the opponent can transition to UP (and never DN), resulting in the play ⇢ = (UK look UP chop)! , i.e., the repetition of (UK look UP chop) forever. Such a play of Aβ chop does not correspond to any play in Achop. This is a well known phenomena of the standard belief-set construction [Sardi ˜na et al., 2006], which our construction overcomes by adding a restriction that removes from consideration plays such as ⇢ (as discussed in Example 5).

The following definition is central. It maps a history h 2 Hist(A) to the corresponding history hβ 2 Hist(Aβ) of belief-states.

Definition 4.1. For h 2 Hist(A) define hβ 2 Hist(Aβ) inductively as follows.

• For s 2 I, define sβ 2 I β to be I \ obs-1 (obs(s)). In words, sβ is the set of initial states the GP game could be in given the observation obs(s).
• If h 2 Hist(A)↪ a 2 Ac↪ s 2 S, then (has)β := hβaB whereB := tr(Last(hβ)↪ a)\obs-1 (obs(s)). Inwords, B is the set of possible states the GP game could be in given the observation sequence obs(has).

In the same way, for ⇡ 2 Ply(A) define ⇡β 2 Ply(A β). Extend the map pointwise to sets of plays P ✓ Ply(A), i.e., define P β := ¶⇡β 2 Ply(A β) ♣ ⇡ 2 P♦. Finally, we give notation to the special case that P = Ply(A): write Im(A) for the set ¶⇡β ♣ ⇡ 2 Ply(A)♦, called the image of A.

By definition, Im(A) ✓ Ply(Aβ). However, the converse is not always true.

Example 4 (continued). There is a play of Aβ chop that is not in Im(Achop), e.g., ⇢ = (uk look up chop)! . Indeed, suppose ⇡β = ⇢ and consider the sequence of counter values of ⇡. Every look action establishes that the current counter value in ⇡ is positive (this is the meaning of the tree being up), but every chop action reduces the current counter value by one.

This contradicts that counter values are always non-negative. Remark 4.2. If A is finitely-branching then Im(A) =

Ply(A β). To see this, let ⇢ be a play in Aβ , and consider the forest whose nodes are the histories h of A such that hβ is a prefix of ⇢. Each tree in the forest is finitely branching (because A is), and at least one tree in this forest is infinite. Thus, by K˝onig's lemma, the tree has an infinite path ⇡. But ⇡ is a play in A and ⇡β = ⇢.

Definition 4.3. For ⇢ 2 Ply(Aβ), say ⇢ = B0a0B1a1 ▷ ▷ ▷ , define obs(⇢) to be the sequence obs(q0)a0obs(q1)a1 ▷ ▷ ▷

where qi 2 Bi for i 2 N0 (this is well defined since, by definition of the state set S β , each Bi is a subset of a unique observation-set).

The classic belief-state construction transforms hA↪Wi into hAβ↪ Wi. Example 3 shows that this transformation may not preserve the agent having a winning strategy if A is infinite. We now define the generalized belief-state construction and the main technical theorem of this work.

Definition 4.4. Let G = hA↪Wi be a GP game. Define Gβ = hAβ↪W↪ Im(A)i, a GP game of perfect information with restriction. The restriction Im(A) ✓ Ply(Aβ) is the image of Ply(A) under the map ⇡ 7! ⇡β .

Theorem 4.5. Let A be a (possibly infinite) arena of imperfect information, Aβ the corresponding belief-state arena of perfect information, and Im(A) ✓ Ply(Aβ) the image of A. Then, for every winning objective W, the agent has a winning observational-strategy in the GP game G = hA↪Wi if and only if the agent has a winning strategy in the GP game Gβ = hAβ↪W↪ Im(A)i. Moreover, if A is finitely-branching then Gβ = hAβ↪ Wi. 5

Proof. The second statement follows from the first statement and Remark 4.2. For the first statement, we first need some facts that immediately follow from Definition 4.1.

1. (h1)β = (h2)β if and only if obs(h1) = obs(h2).
2. For every h 2 Hist(Aβ) that is also a prefix of ⇡β there is a history h0 2 Hist(A) that is also a prefix of ⇡ such that (h0 )β = h. Also, for every h0 2 Hist(A) that is also a prefix of ⇡ there is a history h 2 Hist(Aβ) that is also a prefix of ⇡β such that (h0 )β = h.

Second, there is a natural correspondence between observational strategies of A and strategies of Aβ .

• If σ is a strategy in Aβ then define the strategy !(σ) of A as mapping h 2 Hist(A) to σ(hβ). Now, !(σ) is observational by Fact 1. Also, if ⇡ is consistent with !(σ) then ⇡β is consistent with σ. Indeed, let h be a history that is also a prefix of ⇡β . We need to show that hσ(h) is a prefix of ⇡β . Suppose that σ(h) = a. Take prefix h0 of ⇡ such that (h0 )β = h (Fact 2). Then !(σ)(h0 ) = σ((h0 )β) = σ(h) = a. Since ⇡ is assumed consistent with !(σ), conclude that h0a is a prefix of ⇡. Thus ha is a prefix of ⇡β .
• If σ is an observational strategy in A then define the strategy (σ) of Aβ as mapping h 2 Hist(Aβ) to σ(h0 ) where h0 is any history such that h0β = h. This is welldefined by ( ) and the fact that σ is observational. Also, if ⇢ is consistent with (σ), then every ⇡ with ⇡β = ⇢ (if there are any) is consistent with σ. Indeed, let h0 be a history of ⇡ and take a prefix h of ⇡β such that (h0 )β = h (Fact 2). Then (σ)(h) = σ(h0 ), call this action a. But ⇡β is assumed consistent with (σ), and thus ha is a prefix of ⇡β . Thus h0a is a prefix of ⇡.

We now put everything together and show that the agent has a winning observational-strategy in G iff the agent has a winning strategy in Gβ .

5 The case that A is finite appears in [Raskin et al., 2007].

Suppose σ is a winning strategy in Gβ . Let ⇡ 2 Ply(A) be consistent with the observational strategy !(σ) of G. Then ⇡β 2 Im(A) is consistent with σ. But σ is assumed to be winning, thus obs(⇡β) 2 W. But obs(⇡) = obs(⇡β). Conclude that !(σ) is a winning strategy in G.

Conversely, suppose σ is a winning strategy in G. Let ⇢ 2 Im(A) be consistent with the strategy (σ) of Gβ , and take ⇡ 2 Ply(A) be such that ⇡β = ⇢ (such a ⇡ exists since we assumed ⇢ 2 Im(A)). Then ⇡ is consistent with σ. But σ is assumed to be winning, thus obs(⇡) 2 W. But obs(⇢) = obs(⇡β) = obs(⇡). Conclude that (σ) is a winning strategy in Gβ .

Remark 4.6. The proof of Theorem 4.5 actually shows how to transform strategies between the GP games, i.e., σ 7! !(σ) and σ 7! (σ), and moreover, these transformations are inverses of each other.

We end with our running example:

Example 5 (Continued). Solving Gchop (Figure 1) is equivalent to solving the finite GP game Gβ chop of perfect information, i.e., hAβ chop ↪ W↪ Im(Achop)i, where the arena A is shown in Figure 2. To solve this we should understand the structure of Im(Achop). It is not hard to see that a play ⇢ 2 Ply(Aβ chop ) is in Im(Achop) if and only if it contains only finitely many infixes of the form "UK look UP". This property is expressible in LTL by the formula ≻GF[UK ^ X look ^ XXUP]. Thus we can apply the algorithm for solving finite games of perfect information with LTL objectives (see, e.g., [Pnueli and Rosner, 1989; de Alfaro et al., 2001]) to solve Gβ chop , and thus the original GP game Gchop.

5 Application of the Construction

We now show how to use generalized-planning (GP) games and our generalized belief-state construction to obtain effective planning procedures for sophisticated problems. For the rest of this section we assume Obs is finite (A may be infinite) so that we can consider LTL temporally extended goals over the alphabet Obs. For instance, LTL formulas specify persistent surveillance missions such as "get items from region A, drop items at region B, infinitely often, and always avoid region C".

Definition 5.1. Let ' be an LTL formula over Obs ⇥Ac. For an arena A, define [[']] = ¶⇡ 2 Ply(A) ♣ obs(⇡) ♣= '♦.

The following is immediate from Theorem 4.5 and the fact that solving finite LTL games of perfect information is decidable [Pnueli and Rosner, 1989; de Alfaro et al., 2001]:

Theorem 5.2. Let G = hA↪ [[']]i be a GP game with a finite arena (possibly obtained as the disjoint union of several arenas sharing the same observations), and ' be an LTL winning objective. Then solving G can be reduced to solving the finite GP game Gβ = hAβ↪ [[']]i of perfect information, which is decidable.

Although we defined winning objectives to be observable, one may prefer general winning conditions, i.e., W ✓ S! . In this case, for finite arenas there is a translation from parity-objectives to observable parity-objectives [Chatterjee and Doyen, 2010]; moreover, for reachability objectives, a plan reaches a goal T ✓ S iff it reaches a belief-state B ✓ T [Bertoli et al., 2006].

Next we look a case where the arena is actually infinite. Recently, the AI community has considered games generated by pushdown-automata [Murano and Perelli, 2015; Chen et al., 2016]. However, the games considered are of perfect information and cannot express generalized-planning problems or planning under partial observability. In contrast, our techniques can solve these planning problems on pushdown domains assuming that the stack is not hidden (we remark that if the stack is hidden, then game-solving becomes undecidable [Azhar et al., 2001]):

Theorem 5.3. Let G = hA↪ [[']]i be a GP game with a pushdown-arena with observable stack, and ' is an LTL formula. Then solving G can be reduced to solving Gβ = hAβ↪ [[']]i, a GP game with pushdown-arena with perfect information, which is decidable.

Proof. Let P be a pushdown-automaton with states Q, initial state q0, finite input alphabet Ac, and finite stack alphabet Γ. We call elements of Γ⇤ stacks, and denote the empty stack by ✏. Also, fix an observation function on the states, i.e., f : Q!Obs for some set Obs (we do not introduce notation for the transition function of P). A pushdown-arena AP = (S↪ I↪ Ac↪ tr↪Obs↪ obs) is generated by P as follows: the set of states S is the set of configurations of P, i.e., pairs (q↪ γ) where q 2 Q and γ 2 Γ⇤ is a stack-content of P; the initial state of A is the initial configuration, i.e., I = ¶(q0↪ ✏)♦; the transition function of A is defined as tr((q↪ γ)↪ a) = (q0 ↪ γ 0 ) if P can move in one step from state q and stack content γ to state q0 and stack content γ 0 by consuming the input letter a; the observation function obs maps a configuration (q↪ γ) to f(q) (i.e., this formalizes the statement that the stack is observable). Observe now that: (1) the GP game A is finitely-branching; (2) the GP game Aβ is generated by a pushdown automaton (its states are subsets of Q). Thus we can apply Theorem 4.5 to reduce solving A, a GP-game with imperfect information and pushdown arena, to Aβ , a GP-game with perfect information and pushdown arena. The latter is decidable [Walukiewicz, 2001].

6 Related work in Formal Methods

Games of imperfect information on finite arenas have been studied extensively. Reachability winning-objectives were studied in [Reif, 1984] from a complexity point of view: certain games were shown to be universal in the sense that they are the hardest games of imperfect information, and optimal decision procedures were given. More generally, !-regular winning-objectives were studied in [Raskin et al., 2007], and symbolic algorithms were given (also for the case of randomized strategies).

To solve (imperfect-information) games on infinite arenas one needs a finite-representation of the infinite arena. One canonical way to generate infinite arenas is by parametric means. In this line, [Jacobs and Bloem, 2014] study the synthesis problem for distributed architectures with a parametric number of finite-state components. They leverage results from the Formal Methods literature that say that for certain types of token-passing systems there is a cutoff [Emerson and Namjoshi, 1995], i.e., an upper bound on the number of components one needs to consider in order to synthesize a protocol for any number of components. Another way to generate infinite arenas is as configuration spaces of pushdown automata. These are important in analysis of software because they capture the flow of procedure calls and returns in reactive programs. Module-checking pushdown systems of imperfect information [Bozzelli et al., 2010; Aminof et al., 2013] can be thought of as games in which the environment plays non-deterministic strategies. Although undecidable, by not hiding the stack (cf. Theorem 5.3) decidability of module-checking is regained.

Finally, we note that synthesis of distributed systems has been studied in the Formal Methods literature using the techniques of games, starting with [Pnueli and Rosner, 1990]. Such problems can be cast as multi-player games of imperfect information, and logics such as ATL with knowledge can be used to reason about strategies in these games. However, even for three players, finite arenas, and reachability goals, the synthesis problem (and the corresponding model checking problem for ATLK) is undecidable [Dima and Tiplea, 2011].

7 Critical Evaluation and Conclusions

Although our technique for removing partial observability is sound and complete, it is, necessarily, not algorithmic: indeed, no algorithm can always remove partial observability from computable infinite domains and result in a solvable planning problem (e.g, one with a finite domain). 6

The main avenue for future technical work is to establish natural classes of generalized-planning problems that can be solved algorithmically. We believe the methodology of this paper will be central to this endeavor. Indeed, as we showed in Section 5, we can identify Im(A) in a number of cases. We conjecture that one can do the same for all of the onedimensional planning problems of [Hu and Levesque, 2010; Hu and De Giacomo, 2011].

The framework presented in this paper is non-probabilistic, but extending it with probabilities and utilities associated to agent choices [Kaelbling et al., 1998; LaValle, 2006; Bonet and Geffner, 2009; Geffner and Bonet, 2013] is of great interest. In particular, POMDPs with temporally-extended winning objectives (e.g., LTL, B¨uchi, parity) have been studied for finite domains [Chatterjee et al., 2010]. We leave for future work the problem of dealing with such POMDPs over infinite domains.

Acknowledgments

We thank the reviewers for their constructive comments. This research was partially supported by the Sapienza project "Immersive Cognitive Environments". Aniello Murano was supported in part by GNCS 2016 project: Logica, Automi e Giochi per Sistemi Auto-adattivi. Sasha Rubin is a Marie Curie fellow of the Istituto Nazionale di Alta Matematica.

6 In fact, there is no algorithm solving games of perfect observation on computable domains with reachability objectives.