Generalised planning (GP) refers to the task of synthesising programs that solve families of related planning problems. We introduce a novel, yet simple method for GP: given a set of training problems, for each problem, compute an optimal plan for each goal atom in some order, perform goal regression on the resulting plans, and lift the corresponding outputs to obtain a set of first-order Condition → Actions rules. The rules collectively constitute a generalised plan that can be executed as is or alternatively be used to prune the planning search space. We formalise and prove the conditions under which our method is guaranteed to learn valid generalised plans and state space pruning axioms for search. Experiments demonstrate significant improvements over state-of-the-art (generalised) planners with respect to the 3 metrics of synthesis cost, planning coverage, and solution quality on various classical and numeric planning domains.

Reward machines (RMs) are automata structures that encode (non-Markovian) reward functions for reinforcement learning (RL). RMs can reward any behaviour representable in regular languages and, when paired with RL algorithms that exploit RM structure, have been shown to significantly improve sample efficiency in many domains. In this work, we present pushdown reward machines (pdRMs), an extension of reward machines based on deterministic pushdown automata. pdRMs can recognise and reward temporally extended behaviours representable in deterministic context-free languages, making them more expressive than reward machines. We introduce two variants of pdRM-based policies, one which has access to the entire stack of the pdRM, and one which can only access the top k symbols (for a given constant k) of the stack. We propose a procedure to check when the two kinds of policies (for a given environment, pdRM, and constant k) achieve the same optimal state values. We then provide theoretical results establishing the expressive power of pdRMs, and space complexity results for the proposed learning problems. Lastly, we propose an approach for off-policy RL algorithms that exploits counterfactual experiences with pdRMs. We conclude by providing experimental results showing how agents can be trained to perform tasks representable in deterministic context-free languages using pdRMs.

Grounding language in perception and action is a key challenge when building situated agents that can interact with humans, or other agents, via language. In the past, addressing this challenge has required manually designing the language grounding or curating massive datasets that associate language with the environment. We propose Ground-Compose-Reinforce, an end-to-end, neurosymbolic framework for training RL agents directly from high-level task specifications—without manually designed reward functions or other domain-specific oracles, and without massive datasets. These task specifications take the form of Reward Machines, automata-based representations that capture high-level task structure and are in some cases autoformalizable from natural language. Critically, we show that Reward Machines can be grounded using limited data by exploiting compositionality. Experiments in a custom Meta-World domain with only 350 labelled pretraining trajectories show that our framework faithfully elicits complex behaviours from high-level specifications—including behaviours that never appear in pretraining—while non-compositional approaches fail.

Large Language Model (LLM) agents have demonstrated impressive capabilities for social interaction and are increasingly being deployed in situations where they might engage with both human and artificial agents. These interactions represent a critical frontier for LLM-based agents, yet existing evaluation methods fail to measure how well these capabilities generalize to novel social situations. In this paper, we introduce a method for evaluating the ability of LLM-based agents to cooperate in zero-shot, mixed-motive environments using Concordia, a natural language multi-agent simulation environment. Our method measures general cooperative intelligence by testing an agent's ability to identify and exploit opportunities for mutual gain across diverse partners and contexts. We present empirical results from the NeurIPS 2024 Concordia Contest, where agents were evaluated on their ability to achieve mutual gains across a suite of diverse scenarios ranging from negotiation to collective action problems. Our findings reveal significant gaps between current agent capabilities and the robust generalization required for reliable cooperation, particularly in scenarios demanding persuasion and norm enforcement.

Fair decision making has largely been studied with respect to a single decision. Here we investigate the notion of fairness in the context of sequential decision making where multiple stakeholders can be affected by the outcomes of decisions. We observe that fairness often depends on the history of the sequential decision-making process, and in this sense that it is inherently non-Markovian. We further observe that fairness often needs to be assessed at time points within the process, not just at the end of the process. To advance our understanding of this class of fairness problems, we explore the notion of non-Markovian fairness in the context of sequential decision making. We identify properties of non-Markovian fairness, including notions of long-term, anytime, periodic, and bounded fairness. We explore the interplay between non-Markovian fairness and memory and how memory can support construction of fair policies. Finally, we introduce the FairQCM algorithm, which can automatically augment its training data to improve sample efficiency in the synthesis of fair policies via reinforcement learning.

Reward Machines provide an automaton-inspired structure for specifying instructions, safety constraints, and other temporally extended reward-worthy behaviour. By exposing the underlying structure of a reward function, they enable the decomposition of an RL task, leading to impressive gains in sample efficiency. Although Reward Machines and similar formal specifications have a rich history of application towards sequential decision-making problems, prior frameworks have traditionally ignored ambiguity and uncertainty when interpreting the domain-specific vocabulary forming the building blocks of the reward function. Such uncertainty critically arises in many real-world settings due to factors like partial observability or noisy sensors. In this work, we explore the use of Reward Machines for Deep RL in noisy and uncertain environments. We characterize this problem as a POMDP and propose a suite of RL algorithms that exploit task structure under uncertain interpretation of the domain-specific vocabulary. Through theory and experiments, we expose pitfalls in naive approaches to this problem while simultaneously demonstrating how task structure can be successfully leveraged under noisy interpretations of the vocabulary.

Many important real-world Reinforcement Learning (RL) problems involve partial observability and require policies with memory. Unfortunately, standard deep RL algorithms for partially observable settings typically condition on the full history of interactions and are notoriously difficult to train. We propose a novel deep, partially observable RL algorithm based on modelling belief states — a technique typically used when solving tabular POMDPs, but that has traditionally been difficult to apply to more complex environments. Our approach simplifies policy learning by leveraging state information at training time, that may not be available at deployment time. We do so in two ways: first, we decouple belief state modelling (via unsupervised learning) from policy optimization (via RL); and second, we propose a representation learning approach to capture a compact set of reward-relevant features of the state. Experiments demonstrate the efficacy of our approach on partially observable domains requiring information seeking and long-term memory.

Within the field of automated planning, two areas of study are planning with preferences and epistemic planning. Planning with preferences involves generating plans that optimize for properties of the plan instead of, or in addition to, trying to reach a fixed goal. Epistemic planning allows for planning over the knowledge or belief states of one or more agents for the purpose of achieving epistemic goals (where agents have particular states of knowledge or belief). In this paper we motivate and explore the task of planning with epistemic preferences, proposing a method by which existing automated planning techniques can be combined for this purpose.

In sequential decision making, objective specifications are often underspecified or incomplete, neglecting to take into account potential (negative) side effects. Executing plans without consideration of their side effects can lead to catastrophic outcomes -- a concern recently raised in relation to the safety of AI. In this paper we investigate how to avoid side effects in a symbolic planning setting. We study the notion of minimizing side effects in the context of a planning environment where multiple independent agents co-exist. We define (classes of) negative side effects in terms of their effect on the agency of those other agents. Finally, we show how plans which minimize side effects of different types can be computed via compilations to cost-optimizing symbolic planning, and investigate experimentally.

Text-based games present a unique class of sequential decision making problem in which agents interact with a partially observable, simulated environment via actions and observations conveyed through natural language. Such observations typically include instructions that, in a reinforcement learning (RL) setting, can directly or indirectly guide a player towards completing reward-worthy tasks. In this work, we study the ability of RL agents to follow such instructions. We conduct experiments that show that the performance of state-of-the-art text-based game agents is largely unaffected by the presence or absence of such instructions, and that these agents are typically unable to execute tasks to completion. To further study and address the task of instruction following, we equip RL agents with an internal structured representation of natural language instructions in the form of Linear Temporal Logic (LTL), a formal language that is increasingly used for temporally extended reward specification in RL. Our framework both supports and highlights the benefit of understanding the temporal semantics of instructions and in measuring progress towards achievement of such a temporally extended behaviour. Experiments with 500+ games in TextWorld demonstrate the superior performance of our approach.

Reinforcement learning (RL) methods usually treat reward functions as black boxes. As such, these methods must extensively interact with the environment in order to discover rewards and optimal policies. In most RL applications, however, users have to program the reward function and, hence, there is the opportunity to make the reward function visible – to show the reward function’s code to the RL agent so it can exploit the function’s internal structure to learn optimal policies in a more sample efficient manner. In this paper, we show how to accomplish this idea in two steps. First, we propose reward machines, a type of finite state machine that supports the specification of reward functions while exposing reward function structure. We then describe different methodologies to exploit this structure to support learning, including automated reward shaping, task decomposition, and counterfactual reasoning with off-policy learning. Experiments on tabular and continuous domains, across different tasks and RL agents, show the benefits of exploiting reward structure with respect to sample efficiency and the quality of resultant policies. Finally, by virtue of being a form of finite state machine, reward machines have the expressive power of a regular language and as such support loops, sequences and conditionals, as well as the expression of temporally extended properties typical of linear temporal logic and non-Markovian reward specification.

Agents change their beliefs about the plausibility of various aspects of domain dynamics -- effects of physical actions, results of sensing, and action preconditions -- as a consequence of their interactions with the world. In this paper we propose a way to conveniently represent domain dynamics in the situation calculus to support such belief change. Furthermore, we suggest patterns to follow when writing the axioms that describe the effects of actions, and prove how these patterns can control the extent to which observations change the agent's beliefs about action effects. We also discuss the relation of our work to the AGM postulates for belief revision. Finally, we show how beliefs about domain dynamics can be incorporated into a form of regression rewriting to support reasoning.

In Reinforcement Learning (RL), an agent is guided by the rewards it receives from the reward function. Unfortunately, it may take many interactions with the environment to learn from sparse rewards, and it can be challenging to specify reward functions that reflect complex reward-worthy behavior. We propose using reward machines (RMs), which are automata-based representations that expose reward function structure, as a normal form representation for reward functions. We show how specifications of reward in various formal languages, including LTL and other regular languages, can be automatically translated into RMs, easing the burden of complex reward function specification. We then show how the exposed structure of the reward function can be exploited by tailored q-learning algorithms and automated reward shaping techniques in order to improve the sample efficiency of reinforcement learning methods. Experiments show that these RM-tailored techniques significantly outperform state-of-the-art (deep) RL algorithms, solving problems that otherwise cannot reasonably be solved by existing approaches.

Reward Machines (RMs), originally proposed for specifying problems in Reinforcement Learning (RL), provide a structured, automata-based representation of a reward function that allows an agent to decompose problems into subproblems that can be efficiently learned using off-policy learning. Here we show that RMs can be learned from experience, instead of being specified by the user, and that the resulting problem decomposition can be used to effectively solve partially observable RL problems. We pose the task of learning RMs as a discrete optimization problem where the objective is to find an RM that decomposes the problem into a set of subproblems such that the combination of their optimal memoryless policies is an optimal policy for the original problem. We show the effectiveness of this approach on three partially observable domains, where it significantly outperforms A3C, PPO, and ACER, and discuss its advantages, limitations, and broader potential.

In this paper we propose Reward Machines {—} a type of finite state machine that supports the specification of reward functions while exposing reward function structure to the learner and supporting decomposition. We then present Q-Learning for Reward Machines (QRM), an algorithm which appropriately decomposes the reward machine and uses off-policy q-learning to simultaneously learn subpolicies for the different components. QRM is guaranteed to converge to an optimal policy in the tabular case, in contrast to Hierarchical Reinforcement Learning methods which might converge to suboptimal policies. We demonstrate this behavior experimentally in two discrete domains. We also show how function approximation methods like neural networks can be incorporated into QRM, and that doing so can find better policies more quickly than hierarchical methods in a domain with a continuous state space.