This paper presents a new algorithm based on grammar induction, called AMLSI (Action Model Learning with State machine Interactions), that retro-engineers planning domains described with Planning Problem Description Language (PDDL) by querying a state machine with action sequences and by observing the state transitions. AMLSI takes as input a training set of feasible and infeasible action sequences built from partial observations and returns a PDDL domain. A key issue for domain learning is the ability to plan with the learned domains.It often happens that a small learning error leads to a domain that is unusable for planning. Unlike other algorithms, we show that AMLSI is able to lift this lock by learning domains from partial observations with sufficient accuracy to allow planners to solve new problems.
This paper presents a novel approach to learn PDDL domain called AMLSI (Action Model Learning with State machine Interaction) based on grammar induction. AMLSI learns with no prior knowledge from a training dataset made up of action sequences built by random walks and by observing state transitions. The domain learnt is accurate enough to be used without human proofreading in a planner even with very highly partial and noisy observations. Thus AMLSI takles a key issue for domain learning that is the ability to plan with the learned domains. It often happens that small learning errors lead to domains that are unusable for planning. AMLSI contribution is to learn domains from partial and noisy observations with sufficient accuracy to allow planners to solve new problems. Compared to other approaches, AMLSI uses smaller training datasets and exploits both feasible and infeasible generated action sequences.
This paper presents new approach based on grammar induction called AMLSI (Action Model Learning with State machine Interactions). The AMLSI approach does not require a training dataset of plan traces to work. AMLSI proceeds by trial and error: it queries the system to learn with randomly generated action sequences, and it observes the state transitions of the system, then AMLSI returns a PDDL domain corresponding to the system. A key issue for domain learning is the ability to plan with the learned domains. It often happens that a small learning error leads to a domain that is unusable for planning. Unlike other algorithms, we show that AMLSI is able to lift this lock by learning domains from partial and noisy observations with sufficient accuracy to allow planners to solve new problems.
Hand-encoding PDDL domains is generally considered difficult, tedious and error-prone. The difficulty is even greater when temporal domains have to be encoded. Indeed, actions have a duration and their effects are not instantaneous. In this paper, we present TempAMLSI, an algorithm based on the AMLSI approach to learn temporal domains. TempAMLSI is the first approach able to learn temporal domains with single hard envelopes, and TempAMLSI is the first approach able to deal with both partial and noisy observations. We show experimentally that TempAMLSI learns accurate temporal domains, i.e., temporal domains that can be used without human proofreading to solve new planning problems with different forms of action concurrency.
Hand-encoding PDDL domains is generally considered difficult, tedious and error-prone. Many different machine learning techniques have been proposed to deal with these issues. Recent novel approaches such as AMLSI (Action Model Learning with System Interaction) achieve high level of accuracy, i.e. the learnt domains are accurate enough to solve planning problems without human proofreading. However, in most of the real world applications, training datasets are difficult and costly to acquire and become available gradually over time. To tackle this issue, we present IncrAMLSI, which is a version of AMLSI for incremental training datasets. We show that IncrAMLSI outperforms AMLSI to learn IPC benchmarks and converges with few learning iterations from partial and noisy data.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.
hi@scite.ai
10624 S. Eastern Ave., Ste. A-614
Henderson, NV 89052, USA
Copyright © 2024 scite LLC. All rights reserved.
Made with 💙 for researchers
Part of the Research Solutions Family.