Learning from Demonstration (LfD) allows novice users to teach robots through demonstrations without coding; however, such demonstrations are often suboptimal and can limit robot performance. To better support novices, we investigate the design of a feedback system that enables effective human-robot communication during demonstrations. We first conducted a focus group study (N = 9) to identify effective ways of visualizing key robot information, including joint limits, self-collisions, and manipulability. Guided by these insights, we designed an AR-based real-time feedback system and evaluated it in a between-subjects user study (N = 36) on a 7-DoF collaborative robot. Participants performed two tasks—insertion and pouring—with the second task enabling assessment of participants’ learning across tasks. Results show that real-time feedback reduced demonstration time, increased task completion rate, lowered perceived mental workload, and improved adherence to robot kinematic constraints. These findings demonstrate the effectiveness of the real-time feedback system for intuitive and effective robot teaching.

Learning from Demonstration allows robots to acquire skills from human demonstrations, making them more accessible to a wider range of users. Among different approaches, kinesthetic teaching allows humans to manipulate the robot joints directly, making it effective method for demonstrating constrained tasks. However, robots with kinematic redundancy enable multiple joint configurations to achieve a desired task, which could influence human teaching performance. One one hand, it could make it easier, allowing more freedom to demonstrate the task, but on the other, it also increases the number of joints that needs to be manipulated, potentially affecting cognitive and physical load of the demonstrator. Therefore, it is crucial to investigate how the number of degrees of redundancy (DoR) impact human performance during kinesthetic demonstrations, and then how these demonstrations influence robot performance. We simulated high and low DoR by locking one of the robot joint on a 7-DoF Panda robotic arm. We conducted a within-subject user study (N = 24) with two conditions: unconstrained condition (high DoR) and constrained condition (low DoR). We used a motion capture system to capture participants physical interaction with the robot when demonstrating two tasks: button pressing and cuboid block insertion. The results show that the robot’s DoR significantly affects mental workload, demonstration time, number of failed attempts, and physical interaction with the robot. Likewise, joint constraints significantly influence robot performance, measured by task completion using the learned model. These findings highlight the importance of considering robot DoR during demonstrating constrained tasks, allowing novice users to provide effective demonstrations.

Efficient job scheduling in high-performance computing (HPC) systems necessitates the simultaneous consideration of system-centric objectives, such as maximizing resource utilization, and user-centric objectives, such as minimizing job waiting times. In practice, the relative importance of these objectives is not static, but shifts dynamically in response to fluctuations in workload characteristics and system state. However, existing scheduling frameworks — including both traditional workload managers and reinforcement learning (RL)-based methods — typically rely on fixed policies or fixed reward functions that encode a predetermined combination of objectives. As a result, they lack the flexibility to adjust their scheduling priorities as workload intensity or system conditions change. In this work, we propose MetaPilot, a deep reinforcement learning-based controller designed to enable dynamic adaptation to shifting scheduling objectives. MetaPilot functions as an adaptive decision layer, continuously monitoring real-time system state and workload characteristics in order to select the most appropriate scheduling objective prior to each scheduling event. The selected objective is subsequently optimized by established scheduling frameworks, ensuring seamless integration with existing systems. Comprehensive experiments conducted on real-world HPC workload traces from both CPU-based (Physical) and GPU-enabled (Deeplearn) clusters demonstrate that MetaPilot achieves substantial improvements in both maximum and average job waiting times, as well as in resource utilization across CPU, GPU, and memory dimensions, relative to state-of-the-art baseline schedulers. Notably, MetaPilot reduces maximum waiting time by up to 19% on the Physical partition and achieves best-in-class utilization and balanced waiting times on the Deeplearn partition, thereby demonstrating robust adaptability and effectiveness across heterogeneous HPC workloads.

Agent system mining is a recently introduced type of process mining that takes a bottom-up approach to the data-driven analysis of socio-technical systems that execute business processes in organizations. Instead of the top-down approach used in conventional process mining that studies a system in terms of its global state evolution, agent system mining analyzes the system as if it is composed of autonomous agents, each with its local state and behavior, interacting with other agents and the environment to contribute to the emerging global behavior of the business process. Recently, Agent Miner, the first algorithm for discovering agent systems from event data generated by process-aware information systems, has been proposed. The quality of the agent systems discovered by this algorithm depends on the quality of the agent types (or agents), which are identified from the available information about agent instances in the data. In this paper, we study the suitability and benefits of using methods from the organizational mining subarea of process mining for identifying agent types. The experiments we conduct over real-world datasets confirm the usefulness of such methods for discovering simple, modular, and accurate agent systems. These conclusions are grounded in quality metrics such as the size of discovered models (simplicity), Louvain modularity and the Gini coefficient (modularity), and precision and recall (accuracy). The results confirm the benefits of using organizational mining for identifying agent types when discovering agent systems from event data, leading to the construction of models of superior quality in precision, recall, and simplicity compared to models constructed by state-of-the-art conventional process discovery algorithms.

High-Performance Computing systems frequently undergo architectural changes, such as hardware upgrades or the deployment of new clusters, to meet evolving computational demands. Traditional static schedulers and machine-learning-based approaches struggle to adapt efficiently to these changes, often requiring manual adjustments or extensive retraining. In this paper, we propose a novel approach combining Separate Feature Extraction and Selective Transfer Learning to enable rapid adaptation of Reinforcement Learning-based HPC sched-ulers to new or modified cluster architectures. We evaluate our approach using three real-world HPC clusters, including both CPU and GPU architectures. Our experiments simulate scheduler transitions between these clusters, capturing a wide range of architectural changes and workload variations found in practice. Applied to a state-of-the-art hierarchical RL-based scheduler, our method demonstrates rapid adaptation across diverse system configurations and workloads. In the six transition experiments, the schedulers enhanced with our approach not only quickly outperformed all 21 baseline heuristic methods in terms of jobs' average and maximum waiting time but also achieved performance comparable to schedulers trained from scratch. Notably, even in the most time-consuming transition case, integrating our approach reduced the required training timesteps to just 1.76% of the total timesteps needed for training a scheduler from scratch - representing a 56.8× reduction in training effort. This demonstrates its ability to efficiently adapt existing schedulers to evolving HPC architectures with minimal cost, providing a practical solution for real-world HPC operations.

Large Language Models (LLMs) have shown promise in solving natural language-described planning tasks, but their direct use often leads to inconsistent reasoning and hallucination. While hybrid LLM-symbolic planning pipelines have emerged as a more robust alternative, they typically require extensive expert intervention to refine and validate generated action schemas. It not only limits scalability but also introduces a potential for biased interpretation, as a single expert's interpretation of ambiguous natural language descriptions might not align with the user's actual intent. To address this, we propose a novel approach that constructs an action schema library to generate multiple candidates, accounting for the diverse possible interpretations of natural language descriptions. We further introduce a semantic validation and ranking module that automatically filter and rank these candidates without expert-in-the-loop. The experiments showed our pipeline maintains superiority in planning over the direct LLM planning approach. These findings demonstrate the feasibility of a fully automated end-to-end LLM-symbolic planner that requires no expert intervention, opening up the possibility for a broader audience to engage with AI planning with less prerequisite of domain expertise.

It has been shown recently that physics-based simulation significantly enhances the disassembly capabilities of real-world assemblies with diverse 3D shapes and stringent motion constraints. However, the efficiency suffers when tackling intricate disassembly tasks that require numerous simulations and increased simulation time. In this work, we propose a State-Based Disassembly Planning (SBDP) approach, prioritizing physics-based simulation with translational motion over rotational motion to facilitate autonomy, reducing dependency on human input, while storing intermediate motion states to improve search scalability. We introduce two novel evaluation functions derived from new Directional Blocking Graphs (DBGs) enriched with state information to scale up the search. Our experiments show that SBDP with new evaluation functions and DBGs constraints outperforms the state-of-the-art in disassembly planning in terms of success rate and computational efficiency over benchmark datasets consisting of thousands of physically valid industrial assemblies.

The strong performance of large language models (LLMs) raises extensive discussion on their application to code generation.Recent research suggests continuous program refinements through visible tests to improve code generation accuracy in LLMs.However, these methods suffer from LLMs' inefficiency and limited reasoning capacity.In this work, we propose an LLM programming workflow (LPW) designed to improve both initial code generation and subsequent refinements within a structured two-phase workflow.Specifically, the solution generation phase formulates a solution plan, which is then verified through visible tests to specify the intended natural language solution.Subsequently, the code implementation phase drafts an initial code according to the solution plan and its verification.If the generated code fails the visible tests, the plan verification serves as the intended solution to consistently inform the refinement process for correcting bugs.Compared to state-of-the-art methods across various existing LLMs, LPW significantly improves the Pass@1 accuracy by up to 16.4% on well-established text-tocode generation benchmarks.LPW also sets new state-of-the-art Pass@1 accuracy, achieving 98.2% on HumanEval, 84.8% on MBPP, 59.3% on LiveCode, 62.6% on APPS, and 34.7% on CodeContests, using GPT-4o as the backbone.Our code is publicly available at

We analyse the computational complexity of two types of partial satisfaction planning problems, namely, over-subscription planning and net-benefit planning. Partial satisfaction planning seeks to maximise the rewards, or net-benefit between the rewards and costs, that can be obtained by achieving a set of (candidate) goals, generally under a cost budget. In this paper, we assume that the cost budget is polynomially bounded with respect to the number of atoms. In contrast to previous works on complexity analysis, we view partial satisfaction planning as optimisation tasks instead of decision ones. Accordingly, we classify over-subscription and net-benefit planning problems in classes within the optimisation complexity framework. Let L be the size of the partial satisfaction planning instance, we first show that both optimisation problems are OptP[LO(1)]-complete. We then prove that while over-subscription planning becomes OptP[O(log L)]-complete under uniform goal rewards and uniform action costs, the complexity of net-benefit planning remains unchanged. Our results show that the relative values of rewards and costs may impact the computational complexity of partial satisfaction planning, and that an optimisation perspective can reveal differences that may otherwise be lost in their decision formulations. In addition, an optimisation complexity analysis can provide insights on approximation properties of these problems.

This volume contains the papers accepted for presentation at ICAPS 2025, the Thirty-Fifth International Conference on Automated Planning and Scheduling, to be held in Melbourne, Australia, November 9-14, 2025. The annual ICAPS conference series was formed in 2003 through the merger of two pre-existing biennial conferences, the International Conference on Artificial Intelligence Planning and Scheduling (AIPS) and the European Conference on Planning (ECP). ICAPS continues the traditional high standards of AIPS and ECP as an archival forum for new research in the field of automated planning and scheduling. ICAPS 2025 was co-located with two other events: The International Conference on the Integration of Constraint Programming, Artificial Intelligence, and Operations Research (CPAIOR), and The International Conference on the Principles of Knowledge Representation and Reasoning (KR). Existing research into methods and representations for Automated Planning and Scheduling has drawn heavily from the research conducted by these communities. We believe that the co-location of these conferences with ICAPS can only boost these beneficial relationships.The frontmatter contains a Preface and lists both the ICAPS 2025 Organising Committee and the ICAPS 2025 Program Committee.

The capability of Large Language Models (LLMs) to plan remains a topic of debate. Some critics argue that strategies to boost LLMs' reasoning skills are ineffective in planning tasks, while others report strong outcomes merely from training models on a planning corpus. This paper revisits these claims by developing an end-to-end LLM-based planner and evaluating a range of reasoning-enhancement strategies --- including fine-tuning, Chain-of-Thought (CoT) prompting, and reinforcement learning (RL) --- across multiple dimensions of plan quality: validity, executability, goal satisfiability, and more. Our findings reveal fine-tuning alone is insufficient, especially on out-of-distribution tasks. Strategies like CoT prompting primarily enhance local coherence, yielding higher executability rates --- a necessary prerequisite for validity --- but provide only incremental gains and struggle to ensure global plan validity. Notably, RL guided by a novel Longest Contiguous Common Subsequence reward significantly enhances both executability and validity, particularly on longer-horizon problems. Overall, our research addresses key misconceptions in the LLM-planning literature and underscores reward-driven RL optimization as a promising direction for advancing robust LLM-based planning by jointly improving executability and validity.

This paper presents HPCsim, a modular and extensible simulation framework designed to support research in High-Performance Computing workload management. HPCsim focuses on modeling the full scheduling cycle — including job submission, queuing, job selection, resource allocation, execution, and evaluation — using real traces and structured cluster metadata. Most of the existing simulation tools focus on low-level system behavior but lack the abstractions needed for HPC workload management research. They do not natively support multi-resource job models, batch queuing, or system-level scheduling policies. Moreover, their focus on low-level execution details, such as hardware timing, message passing, and architectural interactions, introduces significant computational overhead, making them inefficient for large-scale experiments in high-level workload management, especially in learning-based scheduling where rapid simulation and frequent policy updates are critical. The framework is built based on Slurm, one of the most widely deployed workload managers, and introduces a structured workload data schema encompassing job traces, node configurations, and switch-based topologies. HPCsim is fully compatible with Gymnasium, enabling integration with reinforcement learning workflows, while also supporting other scheduling strategies. We validate HPCsim through experiments using real traces from a GPU-accelerated HPC cluster and demonstrate its utility for comparing scheduling policies and visualizing topology-sensitive placement behavior. HPCsim is released as an open-source tool to facilitate reproducibility and accelerate research in HPC scheduling.

Process mining (PM)-based goal recognition (GR) techniques, which infer goals or targets based on sequences of observed actions, have shown efficacy in real-world engineering applications. This study explores the applicability of PM-based GR in identifying target poses for users employing powered transhumeral prosthetics. These prosthetics are designed to restore missing anatomical segments below the shoulder, including the hand. In this article, we aim to apply the GR techniques to identify the intended movements of users, enabling the motors on the powered transhumeral prosthesis to execute the desired motions precisely. In this way, a powered transhumeral prosthesis can assist individuals with disabilities in completing movement tasks. PM-based GR techniques were initially designed to infer goals from sequences of observed actions, where discrete event names represent actions. However, the electromyography electrodes and kinematic sensors on powered transhumeral prosthetic devices register sequences of continuous, real-valued data measurements. Therefore, we rely on methods to transform sensor data into discrete events and integrate these methods with the PM-based GR system to develop target pose recognition approaches. Two data transformation approaches are introduced. The first approach relies on the clustering of data measurements collected before the target pose is reached (the clustering approach). The second approach uses the time series of measurements collected while the dynamic user movement to perform linear discriminant analysis (LDA) classification and identify discrete events (the dynamic LDA approach). These methods are evaluated through offline and human-in-the-loop (online) experiments and compared with established techniques, such as static LDA, an LDA classification based on data collected at static target poses, and GR approaches based on neural networks. Real-time human-in-the-loop experiments further validate the effectiveness of the proposed methods, demonstrating that PM-based GR using the dynamic LDA classifier achieves superior F 1 score and balanced accuracy compared to state-of-the-art techniques.

Abstract High-performance computing (HPC) systems face increasing challenges in job scheduling due to the evolving complexity of computational tasks and the growing diversity and heterogeneity of resources. Job allocation is a critical aspect in contemporary HPC systems, due to compute nodes possessing an increased capacity in terms of physical resources and having the capability to execute multiple jobs simultaneously. However, job allocation is often overlooked in existing reinforcement learning (RL)-based schedulers that mainly focus on selecting suitable jobs from the job queue and leave allocation to overly simplistic policies, such as first-available allocation. The bin-packing nature at the node level of modern HPC necessitates more refined and intelligent allocation strategies. This paper introduces HeraSched, a novel hierarchical reinforcement learning (HRL)-based scheduler, adept at intelligent job selection without separate backfilling and heterogeneity-aware allocation, tailored for modern HPC environments. It efficiently manages diverse workloads across CPU and GPU cluster partitions. We evaluate HeraSched using real-world workloads, demonstrating significant improvements in reducing job waiting times and preventing job starvation compared to 27 scheduling combinations. In validation, the best maximum waiting time among compared methods is 78% higher than HeraSched’s result in overloaded CPU partitions. This performance demonstrates HeraSched’s ability to manage intensely stressed workloads and adapt to previously unseen, high-demand scenarios, thereby establishing a new standard in HPC job scheduling.

The Abstraction and Reasoning Corpus (ARC) is a general artificial intelligence benchmark that poses difficulties for pure machine learning methods due to its requirement for fluid intelligence with a focus on reasoning and abstraction. In this work, we introduce an ARC solver, Generalized Planning for Abstract Reasoning (GPAR). It casts an ARC problem as a generalized planning (GP) problem, where a solution is formalized as a planning program with pointers. We express each ARC problem using the standard Planning Domain Definition Language (PDDL) coupled with external functions representing object-centric abstractions. We show how to scale up GP solvers via domain knowledge specific to ARC in the form of restrictions over the actions model, predicates, arguments and valid structure of planning programs. Our experiments demonstrate that GPAR outperforms the state-of-the-art solvers on the object-centric tasks of the ARC, showing the effectiveness of GP and the expressiveness of PDDL to model ARC problems. The challenges provided by the ARC benchmark motivate research to advance existing GP solvers and understand new relations with other planning computational models. Code is available at github.com/you68681/GPAR.

The Model AI Assignments session seeks to gather and dis- seminate the best assignment designs of the Artificial In- telligence (AI) Education community. Recognizing that as- signments form the core of student learning experience, we here present abstracts of five AI assignments from the 2024 session that are easily adoptable, playfully engaging, and flexible for a variety of instructor needs. Assignment spec- ifications and supporting resources may be found at http://modelai.gettysburg.edu.

Count-based exploration methods are widely employed to improve the exploratory behavior of learning agents over sequential decision problems. Meanwhile, Novelty search has achieved success in Classical Planning through recording of the first, but not successive, occurrences of tuples. In order to structure the exploration, however, the number of tuples considered needs to grow exponentially as the search progresses. We propose a new novelty technique, classical count-based novelty, which aims to explore the state space with a constant number of tuples, by leveraging the frequency of each tuple’s appearance in a search tree. We then justify the mechanisms through which lower tuple counts lead the search towards novel tuples. We also introduce algorithmic contributions in the form of a trimmed open list that maintains a constant size by pruning nodes with bad novelty values. These techniques are shown to complement existing novelty heuristics when integrated in a classical solver, achieving competitive results in challenging benchmarks from recent International Planning Competitions. Moreover, adapting our solver as the frontend planner in dual configurations that utilize both memory and time thresholds demonstrates a significant increase in instance coverage, surpassing current state-of-the-art solvers.

Learning from Demonstration (LfD) empowers novice users to teach robots daily life tasks without writing sophisticated code, thereby promoting the democratization of robotics. However, novice users often provide sub-optimal demonstrations, which can potentially impact the robot’s ability to efficiently learn and execute the tasks. Prior research has assessed the quality of demonstrations by evaluating the robot’s task performance; however, the approach remains insufficient to qualify individual demonstrations, leaving the reason for classifying demonstrations as high- or low-quality unknown. Therefore, this simulation-based study aims to quantify the quality of individual demonstration at each step by incorporating motion-related quality features such as manipulability and joint-space jerk. To assess the efficacy of these features, we initially evaluated the given demonstrations—taking into account each quality feature—to rank them from high- to low-quality. Subsequently, we investigated the impact of demonstration’s quality on task performance and the quality of task execution. In this pursuit, we trained a series of LfD models for distinct manipulation tasks: cube lifting and pick-and-place of soda can. Our results illustrate a strong correlation between ranked demonstrations and the quality of task execution. Interestingly, we observed that the quality features have a significant impact on task performance, particularly when the provided demonstrations exhibit diversity in terms of quality. Overall, this analysis enables quantifying the quality of individual demonstrations based on motion-related quality features, thus improving learning from demonstration.

Goal Recognition (GR) is a research problem that studies ways to infer the goal of an intelligent agent based on its observed behavior and knowledge of the environment in which the agent operates. A common assumption of GR is that the environment is static. However, in many real-world scenarios, for example, recognizing customers’ preferences, it is necessary to recognize the goals of multiple agents or multiple goals of a single agent over an extended period. Therefore, it is reasonable to expect the environment to change throughout a series of goal recognition tasks. This paper presents three process mining-based solutions to the problem of adaptive GR in a changing environment implemented as different control strategies of a system for solving standard GR problems. As a standard GR system that gets controlled, we use the system grounded in process mining techniques, as it can adjust its internal GR mechanisms based on data collected while observing the operating agents. We evaluated our control strategies over synthetic and real-world datasets. The synthetic datasets were generated using the extended version of the Goal Recognition Amidst Changing Environments (GRACE) tool. The datasets account for different types of changes and drifts in the environment. The evaluation results demonstrate a trade-off between the GR performance over time and the effort invested in adaptations of the GR mechanisms of the system, showing that few well-planned adaptations can lead to a consistently high GR performance.

Diverse, top-k, and top-quality planning are concerned with the generation of sets of solutions to sequential decision problems. Previously this area has been the domain of classical planners that require a symbolic model of the problem instance. This paper proposes a novel alternative approach that uses Monte Carlo Tree Search (MCTS), enabling application to problems for which only a black-box simulation model is available. We present a procedure for extracting bounded sets of plans from pre-generated search trees in best-first order, and a metric for evaluating the relative quality of paths through a search tree. We demonstrate this approach on a path-planning problem with hidden information, and suggest adaptations to the MCTS algorithm to increase the diversity of generated plans. Our results show that our method can generate diverse and high-quality plan sets in domains where classical planners are not applicable.

A transhumeral prosthesis restores missing anatomical segments below the shoulder, including the hand. Active prostheses utilize real-valued, continuous sensor data to recognize patient target poses, or goals, and proactively move the artificial limb. Previous studies have examined how well the data collected in stationary poses, without considering the time steps, can help discriminate the goals. In this case study paper, we focus on using time series data from surface electromyography electrodes and kinematic sensors to sequentially recognize patients' goals. Our approach involves transforming the data into discrete events and training an existing process mining-based goal recognition system. Results from data collected in a virtual reality setting with ten subjects demonstrate the effectiveness of our proposed goal recognition approach, which achieves significantly better precision and recall than the state-of-the-art machine learning techniques and is less confident when wrong, which is beneficial when approximating smoother movements of prostheses.

It has been shown recently that successful techniques in classical planning, such as goal-oriented heuristics and landmarks, can improve the ability to compute planning programs for generalized planning (GP) problems. In this work, we introduce the notion of action novelty rank, which computes novelty with respect to a planning program, and propose novelty-based generalized planning solvers, which prune a newly generated planning program if its most frequent action repetition is greater than a given bound v, implemented by novelty-based best-first search BFS(v) and its progressive variant PGP(v). Besides, we introduce lifted helpful actions in GP derived from action schemes, and propose new evaluation functions and structural program restrictions to scale up the search. Our experiments show that the new algorithms BFS(v) and PGP(v) outperform the state-of-the-art in GP over the standard generalized planning benchmarks. Practical findings on the above-mentioned methods in generalized planning are briefly discussed.

In multi-agent planning, preserving the agents' privacy has become an increasingly popular research topic. For preserving the agents' privacy, agents jointly compute a plan that achieves mutual goals by keeping certain information private to the individual agents. Unfortunately, this can severely restrict the accuracy of the heuristic functions used while searching for solutions. It has been recently shown that, for centralized planning, blind search algorithms such as width-based search can solve instances of many existing domains in low polynomial time when they feature atomic goals. Moreover, the performance of goal-oriented search can be improved by combining it with width-based search. In this paper, we investigate the usage of width-based search in the context of (decentralised) collaborative multi-agent privacy-preserving planning, addressing the challenges related to the agents' privacy and performance. In particular, we show that width-based search is a very effective approach over several benchmark domains, even when the search is driven by heuristics that roughly estimate the distance from goal states, computed without using the private information of other involved agents. Moreover, we show that the use of width-based techniques can significantly reduce the number of messages transmitted among the agents, better preserving their privacy and improving their performance. An experimental study presented in the paper analyses the effectiveness of our techniques, and compares them with the state-of-the-art of collaborative multi-agent planning.

The problem of goal recognition requests to automatically infer an accurate probability distribution over possible goals an autonomous agent is attempting to achieve in the environment. The state-of-the-art approaches for goal recognition operate under full knowledge of the environment and possible operations the agent can take. This knowledge, however, is often not available in real-world applications. Given historical observations of the agents' behaviors in the environment, we learn skill models that capture how the agents achieved the goals in the past. Next, given fresh observations of an agent, we infer their goals by diagnosing deviations between the observations and all the available skill models. We present a framework that serves as an outline for implementing such data-driven goal recognition systems and its instance system implemented using process mining techniques. The evaluations we conducted using our publicly available implementation confirm that the approach is well-defined, i.e., all system parameters impact its performance, has high accuracy over a wide range of synthetic and real-world domains, which is comparable with the more knowledge-demanding state-of-the-art approaches, and operates fast.

The aim of decentralised multi-agent (DMA) planning is to coordinate a set of agents to jointly achieve a goal while preserving their privacy. Blind search algorithms, such as width-based search, have recently proved to be very effective in the context of centralised automated planning, especially when combined with goal-oriented techniques. In this paper, we discuss a recent line of research in which the usage of width-based search has been extensively studied in the context of DMA planning, addressing the challenges related to the agents' privacy and performance.

We introduce a new algorithm, Regression based Supervised Learning (RSL), for learning per instance Neural Network (NN) defined heuristic functions for classical planning problems. RSL uses regression to select relevant sets of states at a range of different distances from the goal. RSL then formulates a Supervised Learning problem to obtain the parameters that define the NN heuristic, using the selected states labeled with exact or estimated distances to goal states. Our experimental study shows that RSL outperforms, in terms of coverage, previous classical planning NN heuristics functions while requiring a fraction of the training time.

In this paper, we present a novel approach to epistemic planning called planning with perspectives (PWP) that is both more expressive and computationally more efficient than existing state-of-the-art epistemic planning tools. Epistemic planning — planning with knowledge and belief — is essential in many multi-agent and human-agent interaction domains. Most state-of-the-art epistemic planners solve epistemic planning problems by either compiling to propositional classical planning (for example, generating all possible knowledge atoms or compiling epistemic formulae to normal forms); or explicitly encoding Kripke-based semantics. However, these methods become computationally infeasible as problem sizes grow. In this paper, we decompose epistemic planning by delegating reasoning about epistemic formulae to an external solver. We do this by modelling the problem using Functional STRIPS, which is more expressive than standard STRIPS and supports the use of external, black-box functions within action models. Building on recent work that demonstrates the relationship between what an agent ‘sees’ and what it knows, we define the perspective of each agent using an external function, and build a solver for epistemic logic around this. Modellers can customise the perspective function of agents, allowing new epistemic logics to be defined without changing the planner. We ran evaluations on well-known epistemic planning benchmarks to compare an existing state-of-the-art planner, and on new scenarios that demonstrate the expressiveness of the PWP approach. The results show that our PWP planner scales significantly better than the state-of-the-art planner that we compared against, and can express problems more succinctly.

Width-based algorithms search for solutions through a general definition of state novelty. These algorithms have been shown to result in state-of-the-art performance in classical planning, and have been successfully applied to model-based and model-free settings where the dynamics of the problem are given through simulation engines. Width-based algorithms performance is understood theoretically through the notion of planning width, providing polynomial guarantees on their runtime and memory consumption. To facilitate synergies across research communities, this paper summarizes the area of width-based planning, and surveys current and future research directions.

We propose new width-based planning and learning algorithms inspired from a careful analysis of the design decisions made by previous width-based planners. The algorithms are applied over the Atari-2600 games and our best performing algorithm, Novelty guided Critical Path Learning (N-CPL), outperforms the previously introduced width-based planning and learning algorithms $\pi$-IW(1), $\pi$-IW(1)+ and $\pi$-HIW(n, 1). Furthermore, we present a taxonomy of the Atari-2600 games according to some of their defining characteristics. This analysis of the games provides further insight into the behaviour and performance of the algorithms introduced. Namely, for games with large branching factors, and games with sparse meaningful rewards, N-CPL outperforms $\pi$-IW, $\pi$-IW(1)+ and $\pi$-HIW(n, 1).

Width-based planning algorithms have been demonstrated to be competitive with state-of-the-art heuristic search and SAT-based approaches, without requiring access to a model of action effects and preconditions, just access to a black-box simulator. Width-based planners search is guided by a measure of the novelty of states, that requires observations on simulator states to be given as a set of features. This paper proposes agnostic feature mapping mechanisms that define the features online, as exploration progresses and the domain of continuous state variables is revealed. We demonstrate the effectiveness of these features on the OpenAI gym "classical control" suite of benchmarks. We compare our online planners with state-of-the-art deep reinforcement learning algorithms, and show that width-based planners using our features can find policies of the same quality with significantly less computational resources.

In multi-agent planning, agents jointly compute a plan that achieves mutual goals, keeping certain information private to the individual agents. Agents' coordination is achieved through the transmission of messages, but they can be a source of privacy leakage as they can permit a malicious agent to collect information about other agents' search processes and states. In this paper, we investigate the usage of novelty techniques in the context of (decentralised) multi-agent privacy preserving planning, addressing the challenges related to the agents' privacy and performance. In particular, we show that novelty based techniques allow a significant reduction on the number of messages transmitted among agents, increasing their privacy levels and also their performances. An experimental study analyses the effectiveness of our techniques and compares them with the state of-the-art. Finally, we examine the robustness of our approach considering different delays in the messages transmission as would occur in overloaded networks, due for example to massive attacks or critical situations.

In this letter, we report on the use of the UC-Berkeley Pacman Capture the Flag competition as a major assessment component in two large artificial intelligence (AI) courses. Teams of students are to integrate AI techniques to control the behavior of agents in an adversarial version of the well-knownPacmangame. We discuss how the contest was designed, set up, and used for grading, and reflect on some of the outcomes achieved, including student perception. While the contest is very challenging (and, at times, frustrating) for students, it was successful in keeping them engaged. More importantly, a significant majority of students reported enjoying and learning from it. A powerful software tool to automate the competition for large number of teams is made freely available.

It has been shown recently that the performance of greedy best-first search (GBFS) for computing plans that are not necessarily optimal can be improved by adding forms of exploration when reaching heuristic plateaus: from random walks to local GBFS searches. In this work, we address this problem but using structural exploration methods resulting from the ideas of width-based search. Width-based methodsseek novel states, are not goal oriented, and their power has been shown recently in the Atari and GVG-AI video-games. We show first that width-based exploration in GBFS is more effective than GBFS with local GBFS search (GBFS-LS), and then proceed to formulate a simple and general computational framework where standard goal-oriented search (exploitation) and width-based search (structural exploration) are combined to yield a search scheme, best-first width search, that is better than both and which results in classical planning algorithms that outperform the state-of-the-art planners.

We propose an approach to reason about agent planning programs with global information. Agent planning programs can be understood as a network of planning tasks, accommodating long-term goals, non-terminating behaviors, and interactive execution. We provide a technique that relies on reasoning about ``global" dead-ends and that can be incorporated to any planning-based approach to agent planning problems. In doing so, we also introduce the notion of online execution of such planning structures. We provide experimental evidence suggesting the technique yields significant benefits.

Classical planning is concerned with problems where a goal needs to be reached from a known initial state by doing actions with deterministic, known effects. Classical planners, however, deal only with classical problems that can be expressed in declarative planning languages such as STRIPS or PDDL. This prevents their use on problems that are not easy to model declaratively or whose dynamics are given via simulations. Simulators do not provide a declarative representation of actions, but simply return successor states. The question we address in this paper is: can a planner that has access to the structure of states and goals only, approach the performance of planners that also have access to the structure of actions expressed in PDDL? To answer this, we develop domain-independent, black box planning algorithms that completely ignore action structure, and show that they match the performance of state-of-the-art classical planners on the standard planning benchmarks. Effective black box algorithms open up new possibilities for modeling and for expressing control knowledge, which we also illustrate.

The automatic generation of realistic behavior such as tactical intercepts for Unmanned Aerial Vehicles (UAV) in air combat is a challenging problem. State-of-the-art solutions propose hand-crafted algorithms and heuristics whose performance depends heavily on the initial conditions and specific aerodynamic characteristics of the UAVs involved. This demo shows the ability of domain-independent planners, embedded into simulators, to generate on-line, feed-forward, control signals that steer simulated aircraft as best suits the situation.

We have recently shown that classical planning problems can be characterized in terms of a width measure that is bounded and small for most planning benchmark domains when goals are restricted to single atoms. Two simple algorithms have been devised for exploiting this structure: Iterated Width (IW) for achieving atomic goals, that runs in time exponential in the problem width by performing a sequence of pruned breadth first searches, and Serialized IW (SIW) that uses IW in a greedy search for achieving conjunctive goals one goal at a time. While SIW does not use heuristic estimators of any sort, it manages to solve more problems than a Greedy BFS using a heuristic like hadd. Yet, it does not approach the performance of more recent planners like LAMA. In this short paper, we introduce two simple extension to IW and SIW that narrow the performance gap with state-of-the-art planners. The first involves changing the greedy search for achieving the goals one at a time, by a depth-first search that is able to backtrack. The second involves computing a relaxed plan once before going to the next subgoal for making the pruning in the breadth-first procedure less agressive, while keeping IW exponential in the width parameter. The empirical results are interesting as they follow from ideas that are very different from those used in current planners.

Classical planning has been notably successful in synthesizing finite plans to achieve states where propositional goals hold. In the last few years, classical planning has also been extended to incorporate temporally extended goals, expressed in temporal logics such as LTL, to impose restrictions on the state sequences generated by finite plans. In this work, we take the next step and consider the computation of infinite plans for achieving arbitrary LTL goals. We show that infinite plans can also be obtained efficiently by calling a classical planner once over a classical planning encoding that represents and extends the composition of the planning domain and the Büchi automaton representing the goal. This compilation scheme has been implemented and a number of experiments are reported.