Automated Verification of Social Law Robustness in STRIPS

Abstract

Agents operating in a multi-agent environment must consider not just their own actions, but also those of the other agents in the system. Artificial social systems are a well known means for coordinating a set of agents, without requiring centralized planning or online negotiation between agents. Artificial social systems enact a social law which restricts the agents from performing some actions under some circumstances. A good social law prevents the agents from interfering with each other, but does not prevent them from achieving their goals. However, designing good social laws, or even checking whether a proposed social law is good, are hard questions. In this paper, we take a first step towards automating these processes, by formulating criteria for good social laws in a multi-agent planning framework. We then describe an automated technique for verifying if a proposed social law meets these criteria, based on a compilation to classical planning.

On Verifying Linear Execution Strategies in Planning Against Nature

“…In the presence of non-deterministic events, we look for an execution strategy of an agent that, in a nutshell, specifies which action (if any) is applied in a particular state whenever the agent can act. Our execution model is derived from the model used in previous work on social laws (Karpas, Shleyfman, and Tennenholtz 2017) to account for possibly different durations of actions and events even without explicitly specifying those in the action/event description. To resolve potential conflicts between actions and events, we consider in our execution model that sequences of events triggered by nature alternate with the actions of the agent.…”

“…The definition of linear execution strategy (Definition 9) is, however, not constructive as it contains an implicit ambiguity for the agent in deciding when to apply the next action. To make the definition constructive we leverage the concept of wait-for preconditions that represents a fragment of social laws in multi-agent planning (Karpas, Shleyfman, and Tennenholtz 2017). In our case, wait-for preconditions are sets of states that uniquely define when the agent applies its non-switch action and when it switches.…”

“…Recent works (Chrpa, Gemrot, and Pilát 2020;Chrpa, Pilát, and Med 2021) aimed at alleviating "non-deterministic branching" by studying in which conditions a sequential plan is guaranteed to eventually achieve the goal. In contrast to those works, we consider that nature can apply sequences of events at once and provide a formal definition of waitfor preconditions, inspired by the work on social laws in planning (Karpas, Shleyfman, and Tennenholtz 2017), that establish when a given action (in the sequence) can be applied (e.g., the AUV has to wait until the ship leaves before entering the ship's corridor).…”

Learning-Based Synthesis of Social Laws in STRIPS

“…In this work we consider a multi-agent planning (MAP) setting that reasons about multiple non-collaborative, noncommunicating agents. We start with the definition of the modified version of MA-STRIPS (Brafman and Domshlak 2008) introduced by Karpas et al (2017). This modification employs individual goals for agents rather than a single goal for the entire problem.…”

“…We implement this by randomly adding some percentage of the agents' actions to the generated social law. Note that, mistakes that make a social law not-robust can be detected using social law robustness verification (Karpas, Shleyfman, and Tennenholtz 2017) thus, the noise we induce is unidirectional, and affects only allowed actions. Additionally, the number of the allowed actions is usually much smaller than the size of the restricted actions so adding only 5% noise actually double the size of the social law.…”