Rodrigo Toro Icarte

Andrew C. Li, Zizhao Chen, Pashootan Vaezipoor et al.

Natural and formal languages provide an effective mechanism for humans to specify instructions and reward functions. We investigate how to generate policies via RL when reward functions are specified in a symbolic language captured by Reward Machines, an increasingly popular automaton-inspired structure. We are interested in the case where the mapping of environment state to a symbolic (here, Reward Machine) vocabulary -- commonly known as the labelling function -- is uncertain from the perspective of the agent. We formulate the problem of policy learning in Reward Machines with noisy symbolic abstractions as a special class of POMDP optimization problem, and investigate several methods to address the problem, building on existing and new techniques, the latter focused on predicting Reward Machine state, rather than on grounding of individual symbols. We analyze these methods and evaluate them experimentally under varying degrees of uncertainty in the correct interpretation of the symbolic vocabulary. We verify the strength of our approach and the limitation of existing methods via an empirical investigation on both illustrative, toy domains and partially observable, deep RL domains.

Phillip J. K. Christoffersen, Andrew C. Li, Rodrigo Toro Icarte et al.

Many real-world reinforcement learning (RL) problems necessitate learning complex, temporally extended behavior that may only receive reward signal when the behavior is completed. If the reward-worthy behavior is known, it can be specified in terms of a non-Markovian reward function - a function that depends on aspects of the state-action history, rather than just the current state and action. Such reward functions yield sparse rewards, necessitating an inordinate number of experiences to find a policy that captures the reward-worthy pattern of behavior. Recent work has leveraged Knowledge Representation (KR) to provide a symbolic abstraction of aspects of the state that summarize reward-relevant properties of the state-action history and support learning a Markovian decomposition of the problem in terms of an automaton over the KR. Providing such a decomposition has been shown to vastly improve learning rates, especially when coupled with algorithms that exploit automaton structure. Nevertheless, such techniques rely on a priori knowledge of the KR. In this work, we explore how to automatically discover useful state abstractions that support learning automata over the state-action history. The result is an end-to-end algorithm that can learn optimal policies with significantly fewer environment samples than state-of-the-art RL on simple non-Markovian domains.

Andrew C. Li, Pashootan Vaezipoor, Rodrigo Toro Icarte et al.

Deep reinforcement learning has shown promise in discrete domains requiring complex reasoning, including games such as Chess, Go, and Hanabi. However, this type of reasoning is less often observed in long-horizon, continuous domains with high-dimensional observations, where instead RL research has predominantly focused on problems with simple high-level structure (e.g. opening a drawer or moving a robot as fast as possible). Inspired by combinatorially hard optimization problems, we propose a set of robotics tasks which admit many distinct solutions at the high-level, but require reasoning about states and rewards thousands of steps into the future for the best performance. Critically, while RL has traditionally suffered on complex, long-horizon tasks due to sparse rewards, our tasks are carefully designed to be solvable without specialized exploration. Nevertheless, our investigation finds that standard RL methods often neglect long-term effects due to discounting, while general-purpose hierarchical RL approaches struggle unless additional abstract domain knowledge can be exploited.

Juan Hernandez, Diego Fernández, Manuel Cifuentes et al.

The Never Give Up (NGU) algorithm has proven effective in reinforcement learning tasks with sparse rewards by combining episodic novelty and intrinsic motivation. In this work, we extend NGU to multi-agent environments and evaluate its performance in the simple_tag environment from the PettingZoo suite. Compared to a multi-agent DQN baseline, NGU achieves moderately higher returns and more stable learning dynamics. We investigate three design choices: (1) shared replay buffer versus individual replay buffers, (2) sharing episodic novelty among agents using different k thresholds, and (3) using heterogeneous values of the beta parameter. Our results show that NGU with a shared replay buffer yields the best performance and stability, highlighting that the gains come from combining NGU intrinsic exploration with experience sharing. Novelty sharing performs comparably when k = 1 but degrades learning for larger values. Finally, heterogeneous beta values do not improve over a small common value. These findings suggest that NGU can be effectively applied in multi-agent settings when experiences are shared and intrinsic exploration signals are carefully tuned.

Sebastian Hagedorn, Martín Muñoz, Cristian Riveros et al.

Automata learning has many applications in artificial intelligence and software engineering. Central to these applications is the $L^*$ algorithm, introduced by Angluin. The $L^*$ algorithm learns deterministic finite-state automata (DFAs) in polynomial time when provided with a minimally adequate teacher. Unfortunately, the $L^*$ algorithm can only learn DFAs over finite alphabets, which limits its applicability. In this paper, we extend $L^*$ to learn symbolic automata whose transitions use predicates over rational numbers, i.e., over infinite and dense alphabets. Our result makes the $L^*$ algorithm applicable to new settings like (real) RGX, and time series. Furthermore, our proposed algorithm is optimal in the sense that it asks a number of queries to the teacher that is at most linear with respect to the number of transitions, and to the representation size of the predicates.

Cristian Hinostroza, Rodrigo Toro Icarte, Christ Devia et al.

Large language models (LLMs) have revolutionized natural language processing. Understanding their internal mechanisms is crucial for developing more interpretable and optimized architectures. Mechanistic interpretability has led to the development of various methods for assessing layer relevance, with cosine similarity being a widely used tool in the field. On this work, we demonstrate that cosine similarity is a poor proxy for the actual performance degradation caused by layer removal. Our theoretical analysis shows that a layer can exhibit an arbitrarily low cosine similarity score while still being crucial to the model's performance. On the other hand, empirical evidence from a range of LLMs confirms that the correlation between cosine similarity and actual performance degradation is often weak or moderate, leading to misleading interpretations of a transformer's internal mechanisms. We propose a more robust metric for assessing layer relevance: the actual drop in model accuracy resulting from the removal of a layer. Even though it is a computationally costly metric, this approach offers a more accurate picture of layer importance, allowing for more informed pruning strategies and lightweight models. Our findings have significant implications for the development of interpretable LLMs and highlight the need to move beyond cosine similarity in assessing layer relevance.

Nicolas Buzeta, Felipe del Rio, Cristian Hinostroza et al.

Vision Language Models (VLMs) are designed to extend Large Language Models (LLMs) with visual capabilities, yet in this work we observe a surprising phenomenon: VLMs can outperform their underlying LLMs on purely text-only tasks, particularly in long-context information retrieval. To investigate this effect, we build a controlled synthetic retrieval task and find that a transformer trained only on text achieves perfect in-distribution accuracy but fails to generalize out of distribution, while subsequent training on an image-tokenized version of the same task nearly doubles text-only OOD performance. Mechanistic interpretability reveals that visual training changes the model's internal binding strategy: text-only training encourages positional shortcuts, whereas image-based training disrupts them through spatial translation invariance, forcing the model to adopt a more robust symbolic binding mechanism that persists even after text-only examples are reintroduced. We further characterize how binding strategies vary across training regimes, visual encoders, and initializations, and show that analogous shifts occur during pretrained LLM-to-VLM transitions. Our findings suggest that cross-modal training can enhance reasoning and generalization even for tasks grounded in a single modality.

Parand A. Alamdari, Toryn Q. Klassen, Rodrigo Toro Icarte et al.

Pluralistic alignment is concerned with ensuring that an AI system's objectives and behaviors are in harmony with the diversity of human values and perspectives. In this paper we study the notion of pluralistic alignment in the context of agentic AI, and in particular in the context of an agent that is trying to learn a policy in a manner that is mindful of the values and perspective of others in the environment. To this end, we show how being considerate of the future wellbeing and agency of other (human) agents can promote a form of pluralistic alignment.

Felipe del Rio, Alain Raymond-Saez, Daniel Florea et al.

Deep neural networks (DNNs) struggle at systematic generalization (SG). Several studies have evaluated the possibility to promote SG through the proposal of novel architectures, loss functions or training methodologies. Few studies, however, have focused on the role of training data properties in promoting SG. In this work, we investigate the impact of certain data distributional properties, as inductive biases for the SG ability of a multi-modal language model. To this end, we study three different properties. First, data diversity, instantiated as an increase in the possible values a latent property in the training distribution may take. Second, burstiness, where we probabilistically restrict the number of possible values of latent factors on particular inputs during training. Third, latent intervention, where a particular latent factor is altered randomly during training. We find that all three factors significantly enhance SG, with diversity contributing an 89% absolute increase in accuracy in the most affected property. Through a series of experiments, we test various hypotheses to understand why these properties promote SG. Finally, we find that Normalized Mutual Information (NMI) between latent attributes in the training distribution is strongly predictive of out-of-distribution generalization. We find that a mechanism by which lower NMI induces SG is in the geometry of representations. In particular, we find that NMI induces more parallelism in neural representations (i.e., input features coded in parallel neural vectors) of the model, a property related to the capacity of reasoning by analogy.

Rodrigo Toro Icarte, Ethan Waldie, Toryn Q. Klassen et al.

Reinforcement learning (RL) is a central problem in artificial intelligence. This problem consists of defining artificial agents that can learn optimal behaviour by interacting with an environment -- where the optimal behaviour is defined with respect to a reward signal that the agent seeks to maximize. Reward machines (RMs) provide a structured, automata-based representation of a reward function that enables an RL agent to decompose an RL problem into structured subproblems that can be efficiently learned via 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.

Parand Alizadeh Alamdari, Toryn Q. Klassen, Rodrigo Toro Icarte et al.

Recent work in AI safety has highlighted that in sequential decision making, objectives are often underspecified or incomplete. This gives discretion to the acting agent to realize the stated objective in ways that may result in undesirable outcomes. We contend that to learn to act safely, a reinforcement learning (RL) agent should include contemplation of the impact of its actions on the wellbeing and agency of others in the environment, including other acting agents and reactive processes. We endow RL agents with the ability to contemplate such impact by augmenting their reward based on expectation of future return by others in the environment, providing different criteria for characterizing impact. We further endow these agents with the ability to differentially factor this impact into their decision making, manifesting behavior that ranges from self-centred to self-less, as demonstrated by experiments in gridworld environments.

Maayan Shvo, Zhiming Hu, Rodrigo Toro Icarte et al.

Human beings, even small children, quickly become adept at figuring out how to use applications on their mobile devices. Learning to use a new app is often achieved via trial-and-error, accelerated by transfer of knowledge from past experiences with like apps. The prospect of building a smarter smartphone - one that can learn how to achieve tasks using mobile apps - is tantalizing. In this paper we explore the use of Reinforcement Learning (RL) with the goal of advancing this aspiration. We introduce an RL-based framework for learning to accomplish tasks in mobile apps. RL agents are provided with states derived from the underlying representation of on-screen elements, and rewards that are based on progress made in the task. Agents can interact with screen elements by tapping or typing. Our experimental results, over a number of mobile apps, show that RL agents can learn to accomplish multi-step tasks, as well as achieve modest generalization across different apps. More generally, we develop a platform which addresses several engineering challenges to enable an effective RL training environment. Our AppBuddy platform is compatible with OpenAI Gym and includes a suite of mobile apps and benchmark tasks that supports a diversity of RL research in the mobile app setting.

Pashootan Vaezipoor, Andrew Li, Rodrigo Toro Icarte et al.

We address the problem of teaching a deep reinforcement learning (RL) agent to follow instructions in multi-task environments. Instructions are expressed in a well-known formal language -- linear temporal logic (LTL) -- and can specify a diversity of complex, temporally extended behaviours, including conditionals and alternative realizations. Our proposed learning approach exploits the compositional syntax and the semantics of LTL, enabling our RL agent to learn task-conditioned policies that generalize to new instructions, not observed during training. To reduce the overhead of learning LTL semantics, we introduce an environment-agnostic LTL pretraining scheme which improves sample-efficiency in downstream environments. Experiments on discrete and continuous domains target combinatorial task sets of up to $\sim10^{39}$ unique tasks and demonstrate the strength of our approach in learning to solve (unseen) tasks, given LTL instructions.

Maayan Shvo, Andrew C. Li, Rodrigo Toro Icarte et al.

Sequence classification is the task of predicting a class label given a sequence of observations. In many applications such as healthcare monitoring or intrusion detection, early classification is crucial to prompt intervention. In this work, we learn sequence classifiers that favour early classification from an evolving observation trace. While many state-of-the-art sequence classifiers are neural networks, and in particular LSTMs, our classifiers take the form of finite state automata and are learned via discrete optimization. Our automata-based classifiers are interpretable---supporting explanation, counterfactual reasoning, and human-in-the-loop modification---and have strong empirical performance. Experiments over a suite of goal recognition and behaviour classification datasets show our learned automata-based classifiers to have comparable test performance to LSTM-based classifiers, with the added advantage of being interpretable.

Rodrigo Toro Icarte, Toryn Q. Klassen, Richard Valenzano et al.

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.

Rodrigo Toro Icarte, Richard Valenzano, Toryn Q. Klassen et al.

Reinforcement Learning (RL) agents typically learn memoryless policies---policies that only consider the last observation when selecting actions. Learning memoryless policies is efficient and optimal in fully observable environments. However, some form of memory is necessary when RL agents are faced with partial observability. In this paper, we study a lightweight approach to tackle partial observability in RL. We provide the agent with an external memory and additional actions to control what, if anything, is written to the memory. At every step, the current memory state is part of the agent's observation, and the agent selects a tuple of actions: one action that modifies the environment and another that modifies the memory. When the external memory is sufficiently expressive, optimal memoryless policies yield globally optimal solutions. Unfortunately, previous attempts to use external memory in the form of binary memory have produced poor results in practice. Here, we investigate alternative forms of memory in support of learning effective memoryless policies. Our novel forms of memory outperform binary and LSTM-based memory in well-established partially observable domains.

Rodrigo Toro Icarte, Jorge A. Baier, Cristian Ruz et al.

The knowledge representation community has built general-purpose ontologies which contain large amounts of commonsense knowledge over relevant aspects of the world, including useful visual information, e.g.: "a ball is used by a football player", "a tennis player is located at a tennis court". Current state-of-the-art approaches for visual recognition do not exploit these rule-based knowledge sources. Instead, they learn recognition models directly from training examples. In this paper, we study how general-purpose ontologies---specifically, MIT's ConceptNet ontology---can improve the performance of state-of-the-art vision systems. As a testbed, we tackle the problem of sentence-based image retrieval. Our retrieval approach incorporates knowledge from ConceptNet on top of a large pool of object detectors derived from a deep learning technique. In our experiments, we show that ConceptNet can improve performance on a common benchmark dataset. Key to our performance is the use of the ESPGAME dataset to select visually relevant relations from ConceptNet. Consequently, a main conclusion of this work is that general-purpose commonsense ontologies improve performance on visual reasoning tasks when properly filtered to select meaningful visual relations.