Enhancing Numeric-SAM for Learning with Few Observations

A significant challenge in applying planning technology to real-world problems lies in obtaining a planning model that accurately represents the problem's dynamics. Numeric Safe Action Models Learning (N-SAM) is a recently proposed algorithm that addresses this challenge. It is an algorithm designed to learn the preconditions and effects of actions from observations in domains that may involve both discrete and continuous state variables. N-SAM has several attractive properties. It runs in polynomial time and is guaranteed to output an action model that is safe, in the sense that plans generated by it are applicable and will achieve their intended goals. To preserve this safety guarantee, N-SAM must observe a substantial number of examples for each action before it is included in the learned action model. We address this limitation of N-SAM and propose N-SAM*, an enhanced version of N-SAM that always returns an action model where every observed action is applicable at least in some state, even if it was only observed once. N-SAM* does so without compromising the safety of the returned action model. We prove that N-SAM* is optimal in terms of sample complexity compared to any other algorithm that guarantees safety. An empirical study on a set of benchmark domains shows that the action models returned by N-SAM* enable solving significantly more problems compared to the action models returned by N-SAM.

A Formal Metareasoning Model of Concurrent Planning and Execution

Agents that plan and act in the real world must deal with the fact that time passes as they are planning. When timing is tight, there may be insufficient time to complete the search for a plan before it is time to act. By commencing execution before search concludes, one gains time to search by making planning and execution concurrent. However, this incurs the risk of making incorrect action choices, especially if actions are irreversible. This tradeoff between opportunity and risk is the problem addressed in this paper. Our main contribution is to formally define this setting as an abstract metareasoning problem. We find that the abstract problem is intractable. However, we identify special cases that are solvable in polynomial time, develop greedy solution algorithms, and, through tests on instances derived from search problems, find several methods that achieve promising practical performance. This work lays the foundation for a principled time-aware executive that concurrently plans and executes.

Learning a Shield from Catastrophic Action Effects: Never Repeat the Same Mistake

Agents that operate in an unknown environment are bound to make mistakes while learning, including, at least occasionally, some that lead to catastrophic consequences. When humans make catastrophic mistakes, they are expected to learn never to repeat them, such as a toddler who touches a hot stove and immediately learns never to do so again. In this work we consider a novel class of POMDPs, called POMDP with Catastrophic Actions (POMDP-CA) in which pairs of states and actions are labeled as catastrophic. Agents that act in a POMDP-CA do not have a priori knowledge about which (state, action) pairs are catastrophic, thus they are sure to make mistakes when trying to learn any meaningful policy. Rather, their aim is to maximize reward while never repeating mistakes. As a first step of avoiding mistake repetition, we leverage the concept of a shield which prevents agents from executing specific actions from specific states. In particular, we store catastrophic mistakes (unsafe pairs of states and actions) that agents make in a database. Agents are then forbidden to pick actions that appear in the database. This approach is especially useful in a continual learning setting, where groups of agents perform a variety of tasks over time in the same underlying environment. In this setting, a task-agnostic shield can be constructed in a way that stores mistakes made by any agent, such that once one agent in a group makes a mistake the entire group learns to never repeat that mistake. This paper introduces a variant of the PPO algorithm that utilizes this shield, called ShieldPPO, and empirically evaluates it in a controlled environment. Results indicate that ShieldPPO outperforms PPO, as well as baseline methods from the safe reinforcement learning literature, in a range of settings.