Geraud Nangue Tasse

Junkyu Lee, Michael Katz, Don Joven Agravante et al. · ibm-research

Two common approaches to sequential decision-making are AI planning (AIP) and reinforcement learning (RL). Each has strengths and weaknesses. AIP is interpretable, easy to integrate with symbolic knowledge, and often efficient, but requires an up-front logical domain specification and is sensitive to noise; RL only requires specification of rewards and is robust to noise but is sample inefficient and not easily supplied with external knowledge. We propose an integrative approach that combines high-level planning with RL, retaining interpretability, transfer, and efficiency, while allowing for robust learning of the lower-level plan actions. Our approach defines options in hierarchical reinforcement learning (HRL) from AIP operators by establishing a correspondence between the state transition model of AI planning problem and the abstract state transition system of a Markov Decision Process (MDP). Options are learned by adding intrinsic rewards to encourage consistency between the MDP and AIP transition models. We demonstrate the benefit of our integrated approach by comparing the performance of RL and HRL algorithms in both MiniGrid and N-rooms environments, showing the advantage of our method over the existing ones.

Geraud Nangue Tasse, Matthew Riemer, Benjamin Rosman et al.

Recent success in developing increasingly general purpose agents based on sequence models has led to increased focus on the problem of deploying computationally limited agents within the vastly more complex real-world. A key challenge experienced in these more realistic domains is highly non-Markovian dependencies with respect to the agent's observations, which are less common in small controlled domains. The predominant approach for dealing with this in the literature is to stack together a window of the most recent observations (Frame Stacking), but this window size must grow with the degree of non-Markovian dependencies, which results in prohibitive computational and memory requirements for both action inference and learning. In this paper, we are motivated by the insight that in many environments that are highly non-Markovian with respect to time, the environment only causally depends on a relatively small number of observations over that time-scale. A natural direction would then be to consider meta-algorithms that maintain relatively small adaptive stacks of memories such that it is possible to express highly non-Markovian dependencies with respect to time while considering fewer observations at each step and thus experience substantial savings in both compute and memory requirements. Hence, we propose a meta-algorithm (Adaptive Stacking) for achieving exactly that with convergence guarantees and quantify the reduced computation and memory constraints for MLP, LSTM, and Transformer-based agents. Our experiments utilize popular memory tasks, which give us control over the degree of non-Markovian dependencies. This allows us to demonstrate that an appropriate meta-algorithm can learn the removal of memories not predictive of future rewards without excessive removal of important experiences. Code: https://github.com/geraudnt/adaptive-stacking

Michael Beukman, Branden Ingram, Geraud Nangue Tasse et al.

Reinforcement learning (RL) has progressed substantially over the past decade, with much of this progress being driven by benchmarks. Many benchmarks are focused on video or board games, and a large number of robotics benchmarks lack diversity and real-world applicability. In this paper, we aim to simplify the process of applying reinforcement learning in the 3D simulation league of Robocup, a robotic football competition. To this end, we introduce a Robocup-based RL environment based on the open source rcssserver3d soccer server, simple pre-defined tasks, and integration with a popular RL library, Stable Baselines 3. Our environment enables the creation of high-dimensional continuous control tasks within a robotics football simulation. In each task, an RL agent controls a simulated Nao robot, and can interact with the ball or other agents. We open-source our environment and training code at https://github.com/Michael-Beukman/RobocupGym.

Geraud Nangue Tasse, Benjamin Rosman, Steven James

We propose world value functions (WVFs), a type of goal-oriented general value function that represents how to solve not just a given task, but any other goal-reaching task in an agent's environment. This is achieved by equipping an agent with an internal goal space defined as all the world states where it experiences a terminal transition. The agent can then modify the standard task rewards to define its own reward function, which provably drives it to learn how to achieve all reachable internal goals, and the value of doing so in the current task. We demonstrate two key benefits of WVFs in the context of learning and planning. In particular, given a learned WVF, an agent can compute the optimal policy in a new task by simply estimating the task's reward function. Furthermore, we show that WVFs also implicitly encode the transition dynamics of the environment, and so can be used to perform planning. Experimental results show that WVFs can be learned faster than regular value functions, while their ability to infer the environment's dynamics can be used to integrate learning and planning methods to further improve sample efficiency.

Geraud Nangue Tasse, Steven James, Benjamin Rosman

An open problem in artificial intelligence is how to learn and represent knowledge that is sufficient for a general agent that needs to solve multiple tasks in a given world. In this work we propose world value functions (WVFs), which are a type of general value function with mastery of the world - they represent not only how to solve a given task, but also how to solve any other goal-reaching task. To achieve this, we equip the agent with an internal goal space defined as all the world states where it experiences a terminal transition - a task outcome. The agent can then modify task rewards to define its own reward function, which provably drives it to learn how to achieve all achievable internal goals, and the value of doing so in the current task. We demonstrate a number of benefits of WVFs. When the agent's internal goal space is the entire state space, we demonstrate that the transition function can be inferred from the learned WVF, which allows the agent to plan using learned value functions. Additionally, we show that for tasks in the same world, a pretrained agent that has learned any WVF can then infer the policy and value function for any new task directly from its rewards. Finally, an important property for long-lived agents is the ability to reuse existing knowledge to solve new tasks. Using WVFs as the knowledge representation for learned tasks, we show that an agent is able to solve their logical combination zero-shot, resulting in a combinatorially increasing number of skills throughout their lifetime.

Geraud Nangue Tasse, Devon Jarvis, Steven James et al.

It is desirable for an agent to be able to solve a rich variety of problems that can be specified through language in the same environment. A popular approach towards obtaining such agents is to reuse skills learned in prior tasks to generalise compositionally to new ones. However, this is a challenging problem due to the curse of dimensionality induced by the combinatorially large number of ways high-level goals can be combined both logically and temporally in language. To address this problem, we propose a framework where an agent first learns a sufficient set of skill primitives to achieve all high-level goals in its environment. The agent can then flexibly compose them both logically and temporally to provably achieve temporal logic specifications in any regular language, such as regular fragments of linear temporal logic. This provides the agent with the ability to map from complex temporal logic task specifications to near-optimal behaviours zero-shot. We demonstrate this experimentally in a tabular setting, as well as in a high-dimensional video game and continuous control environment. Finally, we also demonstrate that the performance of skill machines can be improved with regular off-policy reinforcement learning algorithms when optimal behaviours are desired.

Damion Harvey, Geraud Nangue Tasse, Branden Ingram et al.

We consider the problem of unsupervised skill segmentation and hierarchical structure discovery in reinforcement learning. While recent approaches have sought to segment trajectories into reusable skills or options, most rely on action labels, rewards, or handcrafted annotations, limiting their applicability. We propose a method that segments unlabelled trajectories into skills and induces a hierarchical structure over them using a grammar-based approach. The resulting hierarchy captures both low-level behaviours and their composition into higher-level skills. We evaluate our approach in high-dimensional, pixel-based environments, including Craftax and the full, unmodified version of Minecraft. Using metrics for skill segmentation, reuse, and hierarchy quality, we find that our method consistently produces more structured and semantically meaningful hierarchies than existing baselines. Furthermore, as a proof of concept for utility, we demonstrate that these discovered hierarchies accelerate and stabilise learning on downstream reinforcement learning tasks.

Simon Rosen, Siddarth Singh, Ebenezer Gelo et al.

Evaluating moral alignment in agents navigating conflicting, hierarchically structured human norms is a critical challenge at the intersection of AI safety, moral philosophy, and cognitive science. We introduce Morality Chains, a novel formalism for representing moral norms as ordered deontic constraints, and MoralityGym, a benchmark of 98 ethical-dilemma problems presented as trolley-dilemma-style Gymnasium environments. By decoupling task-solving from moral evaluation and introducing a novel Morality Metric, MoralityGym allows the integration of insights from psychology and philosophy into the evaluation of norm-sensitive reasoning. Baseline results with Safe RL methods reveal key limitations, underscoring the need for more principled approaches to ethical decision-making. This work provides a foundation for developing AI systems that behave more reliably, transparently, and ethically in complex real-world contexts.

Tristan Bester, Benjamin Rosman, Steven James et al.

We present counting reward automata-a finite state machine variant capable of modelling any reward function expressible as a formal language. Unlike previous approaches, which are limited to the expression of tasks as regular languages, our framework allows for tasks described by unrestricted grammars. We prove that an agent equipped with such an abstract machine is able to solve a larger set of tasks than those utilising current approaches. We show that this increase in expressive power does not come at the cost of increased automaton complexity. A selection of learning algorithms are presented which exploit automaton structure to improve sample efficiency. We show that the state machines required in our formulation can be specified from natural language task descriptions using large language models. Empirical results demonstrate that our method outperforms competing approaches in terms of sample efficiency, automaton complexity, and task completion.

Vanya Cohen, Geraud Nangue Tasse, Nakul Gopalan et al.

Combining reinforcement learning with language grounding is challenging as the agent needs to explore the environment while simultaneously learning multiple language-conditioned tasks. To address this, we introduce a novel method: the compositionally-enabled reinforcement learning language agent (CERLLA). Our method reduces the sample complexity of tasks specified with language by leveraging compositional policy representations and a semantic parser trained using reinforcement learning and in-context learning. We evaluate our approach in an environment requiring function approximation and demonstrate compositional generalization to novel tasks. Our method significantly outperforms the previous best non-compositional baseline in terms of sample complexity on 162 tasks designed to test compositional generalization. Our model attains a higher success rate and learns in fewer steps than the non-compositional baseline. It reaches a success rate equal to an oracle policy's upper-bound performance of 92%. With the same number of environment steps, the baseline only reaches a success rate of 80%.

Geraud Nangue Tasse, Tamlin Love, Mark Nemecek et al.

An important problem in reinforcement learning is designing agents that learn to solve tasks safely in an environment. A common solution is for a human expert to define either a penalty in the reward function or a cost to be minimised when reaching unsafe states. However, this is non-trivial, since too small a penalty may lead to agents that reach unsafe states, while too large a penalty increases the time to convergence. Additionally, the difficulty in designing reward or cost functions can increase with the complexity of the problem. Hence, for a given environment with a given set of unsafe states, we are interested in finding the upper bound of rewards at unsafe states whose optimal policies minimise the probability of reaching those unsafe states, irrespective of task rewards. We refer to this exact upper bound as the "Minmax penalty", and show that it can be obtained by taking into account both the controllability and diameter of an environment. We provide a simple practical model-free algorithm for an agent to learn this Minmax penalty while learning the task policy, and demonstrate that using it leads to agents that learn safe policies in high-dimensional continuous control environments.

Vanya Cohen, Geraud Nangue Tasse, Nakul Gopalan et al.

We propose a framework that learns to execute natural language instructions in an environment consisting of goal-reaching tasks that share components of their task descriptions. Our approach leverages the compositionality of both value functions and language, with the aim of reducing the sample complexity of learning novel tasks. First, we train a reinforcement learning agent to learn value functions that can be subsequently composed through a Boolean algebra to solve novel tasks. Second, we fine-tune a seq2seq model pretrained on web-scale corpora to map language to logical expressions that specify the required value function compositions. Evaluating our agent in the BabyAI domain, we observe a decrease of 86% in the number of training steps needed to learn a second task after mastering a single task. Results from ablation studies further indicate that it is the combination of compositional value functions and language representations that allows the agent to quickly generalize to new tasks.

Geraud Nangue Tasse, Steven James, Benjamin Rosman

The ability to compose learned skills to solve new tasks is an important property of lifelong-learning agents. In this work, we formalise the logical composition of tasks as a Boolean algebra. This allows us to formulate new tasks in terms of the negation, disjunction and conjunction of a set of base tasks. We then show that by learning goal-oriented value functions and restricting the transition dynamics of the tasks, an agent can solve these new tasks with no further learning. We prove that by composing these value functions in specific ways, we immediately recover the optimal policies for all tasks expressible under the Boolean algebra. We verify our approach in two domains---including a high-dimensional video game environment requiring function approximation---where an agent first learns a set of base skills, and then composes them to solve a super-exponential number of new tasks.