Keychain Problem in Sequential Optimization

• Keychain Problem is a sequential decision-making framework that selects actions (keys) from available subsets (keychains) based on Bayesian priors to maximize successful unlocks.
• It employs combinatorial optimization techniques such as maximum-weight bipartite matching and laminar matching to optimize expected rewards under uncertainty.
• The framework extends to scenarios with multiple correct keys and adversarial priors, offering approximation algorithms and complexity insights for online matching applications.

The Keychain Problem refers to a family of sequential decision-making scenarios in which an agent (often dubbed the "locksmith") chooses actions from available subsets (keychains) to maximize expected payoff, typically the number of successfully opened locks. The challenge arises because, at each stage, the agent sees only a subset of actions and must act based on uncertain knowledge (a Bayesian prior) about which actions are actually effective. Opportunity cost is minimized when the agent maximizes the rounds in which a correct key is present and selected. The problem generalizes to important domains in combinatorial optimization, policy design, and online matching.

1. Formal Definition and Fundamental Structure

In the canonical Keychain Problem, the agent navigates a sequence of rounds; in each round, a subset of actions ("keychain") is presented, and the agent selects one action ("key") to attempt. There is at least one "correct key"—a hidden variable drawn from a known Bayesian prior—which succeeds in opening the lock when selected. The objective is to maximize the total number of successful rounds, equivalently, minimize the opportunity cost: the expected number of rounds where a correct key was present but not chosen.

Formally, for a sequence of $n$ keychains $\mathcal{C}_1, \ldots, \mathcal{C}_n$, each $\mathcal{C}_t \subseteq [m]$ is a subset of the $m$ possible keys. The agent has prior $P$ over which keys are correct. At each round $t$, the policy maps $\mathcal{C}_1, \ldots, \mathcal{C}_t$ and prior knowledge to a key $k \in \mathcal{C}_t$.

2. Algorithmic Solutions: Single and Multiple Correct Keys

When the problem is restricted to a fixed, known order of keychains and a unique correct key, the optimal solution is "exploitative": upon identifying the correct key, the policy always selects it whenever it appears in subsequent chains. Any non-exploitative policy is strictly dominated in expected reward, as proved via exchange arguments.

Under these assumptions, the selection problem is equivalent to a maximum-weight bipartite matching. The bipartite graph consists of keys (nodes on the left) and ordered keychains (nodes on the right); edge weights correspond to the expected future appearances in the chains given the prior. Hence, the Bayes-optimal policy can be computed in polynomial time using classical matching algorithms.

With multiple correct keys, the situation becomes formally intractable: optimal exploitation may lead to sub-optimal information gain about which keys work, increasing opportunity cost. Explicit reductions from Vertex Cover show that the general problem is APX-hard; for instance, no polynomial-time algorithm can approximate the optimum within a factor better than $(4063/4064+\varepsilon)$ unless P=NP.

3. Uncertainty in Keychain Order: Scenarios and Combinatorial Auctions

A major extension is handling random or uncertain orderings of keychains—termed "Probabilistic Scenarios". The prior here comprises both a distribution over which keys are correct and over which keychain sequences ("scenarios") will occur. The paper demonstrates that it suffices to focus on deterministic policies over "information sets", defined as prefixes of observed keychains.

This scenario is reduced to the Maximum Weight Laminar Matching (MWLM) problem: a bipartite graph where each right node (information set) is tagged by the set of scenarios—and these scenario sets form a laminar family. Each edge weight is the product of a scenario's probability and the expected future reward if a key is selected.

Further, MWLM is reduced to a combinatorial auction with XOS valuations, in which each key is a bidder and each information set an item; the bidders' valuations are encoded as antichain functions. Established approximation algorithms for XOS combinatorial auctions produce policies achieving a $(1-1/e)$-approximation of the optimal expected reward.

A matching hardness result—via reduction from Max-(3,2B)-SAT—shows that no algorithm can exceed the $(4063/4064)$-approximation threshold for general scenarios.

4. Extensions: Sampling and Adversarial Priors

When the prior distribution is only accessible as a black-box sampling oracle, the value of each decision is estimated via repeated sampling and concentration bounds (union/Hoeffding), yielding a high-probability approximation within a chosen additive error. The policy then computes a $(1-1/e)-\varepsilon$ approximation to the online reward.

The paper also introduces results in adversarial uncertainty. When the distribution is only known to belong to a convex set, an online learning algorithm such as Follow-the-Regularized-Leader (with entropy regularization) attains near-minimax regret guarantees: performance is within additive $\varepsilon$ of the best achievable policy against the worst-case prior.

5. Keychain Order Selection and Complexity

Another variant empowers the agent to select the order in which keychains are processed, impacting the total expected reward. The process becomes a joint optimization over both orderings and per-keychain decisions.

A simple algorithm is proved to achieve a $1/2$-approximation: compute the best assignment for an arbitrary fixed order and its reverse, and select the higher-reward policy. By an averaging argument, one must attain at least half the optimal reward in one of these cases.

However, computing the Bayes-optimal policy (in terms of sequence ordering and key assignment) is NP-hard, explicitly via reduction from the Upper Triangular Matrix Permutation problem.

6. Reductions to Online Bipartite Matching and Generalizations

The combinatorial techniques developed for the Keychain Problem directly yield new policies and approximation guarantees for variants of online bipartite matching and b-matching. Reductions show that stochastic models for online bipartite matching can be analyzed as laminar matching problems and, with combinatorial auction tools, enable $(1-1/e)$-approximate strategies. Associated hardness theorems for these problems follow from the original keychain reductions.

7. Synthesis and Applications

The Keychain Problem encapsulates a broad class of sequential decision-making, unifying ideas from dynamic programming, combinatorial optimization, and policy design. The reductions and approximation results present a systematic framework for approaching stochastic and adversarial versions of online resource selection. Practical applications include online matching algorithms, combinatorial auction design, and policy synthesis for systems constrained by uncertainty in availability and efficacy of actions.

The theoretical depth—exact algorithms for single-correct-key, hardness and approximation results for multi-key and scenario versions, and algorithmic frameworks for adversarial settings—positions the Keychain Problem as a general abstraction for opportunity cost minimization in sequential selection under uncertainty, with transfer to online matching and auction domains (Vuong et al., 7 Sep 2025).