wiki

Responsibility in a Multi-Value Strategic Setting

4 min read

Responsibility in a Multi-Value Strategic Setting

Summary

Parker, Grandi, and Lorini extend responsibility attribution and anticipation from single outcomes to multi-value multi-agent settings. The paper models agents in finite multiagent transition systems, evaluates finite histories against prioritized LTLf value bases, and asks which strategy choices make an agent responsible for avoidable value satisfactions or violations.

The main wiki takeaway is that multi-value responsibility can be made useful for ex ante strategy selection: minimizing passive responsibility matches symbolic regret minimization, while minimizing inexcusable passive responsibility matches non-dominated strategy choice. See Responsibility Anticipation and Multi-Value Responsibility Anticipation for the reusable formal payload.

Key Claims

  • Responsibility for a set of values cannot be reduced to responsibility for each value separately, because the alternative strategy must make the relevant values jointly different in one counterfactual history.
  • A “passive responsibility” set defined by comparing the actual history with a counterfactual strategy satisfies consistency and completeness requirements for multi-value responsibility.
  • Adding weak excuses yields “inexcusable passive responsibility”: an agent is not inexcusably responsible through an alternative strategy if the actual strategy can be rationally defended by some possible behavior of the other agents.
  • Responsibility anticipation should use the worst responsibility set from a single possible history, rather than unioning everything the agent might be responsible for across histories.
  • Passive responsibility minimization is equivalent to symbolic regret minimization. Inexcusable passive responsibility minimization is equivalent to selecting non-dominated strategies, and at least one strategy can satisfy both criteria in the finite setting.

Methods / Formalism

  • The model uses a finite set of agents, finite action names, joint actions, and a multiagent transition system G = (S, tau). Strategies map finite histories to actions.
  • Values are represented as a prioritized value base Omega = (Omega_1, ..., Omega_m) of LTLf formulas. The paper compares satisfied value sets with a lexicographic score over priority levels.
  • A moral action system is nabla = (G, s_0, k, Omega), where k is the finite horizon and Play(sigma, nabla) is the resulting finite history.
  • The central responsibility-set schema compares the actual play with a counterfactual unilateral strategy:
  • The formal results relate the responsibility notions to standard strategy criteria: passive responsibility to regret minimization, and inexcusable passive responsibility to non-dominance.
  • Scope assumptions are strong: the transition system, initial state, horizon, and shared value base are observable; agents do not communicate; and the paper does not assign probabilities over other agents’ strategies.

Evidence / Experiments

  • This is a formal paper with toy multi-agent tables and proof sketches, not an empirical study.
  • The examples use a cleaning-robot setting to show why single-value or union-style responsibility attribution can give misleading multi-value judgments.

Connections

  • Responsibility Anticipation captures the paper’s reusable conceptual contribution: using anticipated responsibility as a strategy-selection criterion.
  • Multi-Value Responsibility Anticipation records the value-base comparison, responsibility-set definitions, weak excuses, and the regret/non-dominance equivalence results.
  • Strategic Reasoning is the primary area because the paper evaluates strategies under uncertainty about other agents’ strategies.
  • Decision Theory is the relevant foundation because the results reinterpret regret and dominance criteria through responsibility.
  • Linear Temporal Logic supplies the finite-trace value language; the paper uses LTLf values rather than reward machines or ATL formulas.
  • Actual Causality is adjacent but not identical: this paper focuses on responsibility for avoidable outcomes under strategy deviations, especially ex ante anticipation, rather than token-cause diagnosis in a fixed causal model.
  • Ghasemi2025 - Toward Virtuous Reinforcement Learning is a nearby ethics-in-RL source, but Parker et al. give a much more explicit formal strategy-comparison model.

Open Questions

  • How computationally tractable are the proposed responsibility notions once strategies are represented compactly rather than extensionally?
  • What changes when agents have partial observability, heterogeneous value bases, communication, or probabilistic beliefs about other agents?
  • Can the symbolic-regret explanation benefit survive when value priorities must be learned or elicited from humans?
  • How should responsibility anticipation connect to responsibility for risk and to causal responsibility under explicit structural models?

Citation

Parker, T., Grandi, U., and Lorini, E. (2024). Responsibility in a Multi-Value Strategic Setting. arXiv:2410.17229v2.

Graph View

Backlinks