Francisco S. Melo

Diogo S. Carvalho, Pedro A. Santos, Francisco S. Melo

We study the convergence of $Q$-learning with linear function approximation. Our key contribution is the introduction of a novel multi-Bellman operator that extends the traditional Bellman operator. By exploring the properties of this operator, we identify conditions under which the projected multi-Bellman operator becomes contractive, providing improved fixed-point guarantees compared to the Bellman operator. To leverage these insights, we propose the multi $Q$-learning algorithm with linear function approximation. We demonstrate that this algorithm converges to the fixed-point of the projected multi-Bellman operator, yielding solutions of arbitrary accuracy. Finally, we validate our approach by applying it to well-known environments, showcasing the effectiveness and applicability of our findings.

Pedro P. Santos, Fábio Vital, Alberto Sardinha et al.

We study general-utility Markov decision processes (GUMDPs) with risk-aware objectives. In this framework, an agent aims to optimize a risk measure of the distribution of objective values, where the objective function depends on the frequency of visitation of states induced by the agent's policy. First, we motivate, propose, and formalize risk-aware GUMDPs, which enable agents and decision makers to trade off expected performance by risk aversion while benefiting from the rich set of objectives that can be cast under the framework of GUMDPs. We focus our attention on the entropic risk measure (ERM). Second, we show how we can solve risk-aware GUMDPs with ERM objectives by resorting to online planning techniques. In particular, we propose an approach based on Monte Carlo Tree Search (MCTS) to provably solve risk-aware GUMDPs up to any desired accuracy. Third, we provide a set of experimental results showcasing that our approach is successful when optimizing for a spectrum of risk-aware behaviors in the context of GUMDPs under diverse tasks (standard MDPs, maximum state entropy exploration, imitation learning, and multi-objective MDPs).

Miguel Faria, Francisco S. Melo, Ana Paiva

In this paper we investigate the notion of legibility in sequential decision tasks under uncertainty. Previous works that extend legibility to scenarios beyond robot motion either focus on deterministic settings or are computationally too expensive. Our proposed approach, dubbed PoL-MDP, is able to handle uncertainty while remaining computationally tractable. We establish the advantages of our approach against state-of-the-art approaches in several simulated scenarios of different complexity. We also showcase the use of our legible policies as demonstrations for an inverse reinforcement learning agent, establishing their superiority against the commonly used demonstrations based on the optimal policy. Finally, we assess the legibility of our computed policies through a user study where people are asked to infer the goal of a mobile robot following a legible policy by observing its actions.

Rustam Zayanov, Francisco S. Melo, Manuel Lopes

We study the problem of teaching via demonstrations in sequential decision-making tasks. In particular, we focus on the situation when the teacher has no access to the learner's model and policy, and the feedback from the learner is limited to trajectories that start from states selected by the teacher. The necessity to select the starting states and infer the learner's policy creates an opportunity for using the methods of inverse reinforcement learning and active learning by the teacher. In this work, we formalize the teaching process with limited feedback and propose an algorithm that solves this teaching problem. The algorithm uses a modified version of the active value-at-risk method to select the starting states, a modified maximum causal entropy algorithm to infer the policy, and the difficulty score ratio method to choose the teaching demonstrations. We test the algorithm in a synthetic car driving environment and conclude that the proposed algorithm is an effective solution when the learner's feedback is limited.

Pedro P. Santos, Diogo S. Carvalho, Miguel Vasco et al.

We introduce hybrid execution in multi-agent reinforcement learning (MARL), a new paradigm in which agents aim to successfully complete cooperative tasks with arbitrary communication levels at execution time by taking advantage of information-sharing among the agents. Under hybrid execution, the communication level can range from a setting in which no communication is allowed between agents (fully decentralized), to a setting featuring full communication (fully centralized), but the agents do not know beforehand which communication level they will encounter at execution time. To formalize our setting, we define a new class of multi-agent partially observable Markov decision processes (POMDPs) that we name hybrid-POMDPs, which explicitly model a communication process between the agents. We contribute MARO, an approach that makes use of an auto-regressive predictive model, trained in a centralized manner, to estimate missing agents' observations at execution time. We evaluate MARO on standard scenarios and extensions of previous benchmarks tailored to emphasize the negative impact of partial observability in MARL. Experimental results show that our method consistently outperforms relevant baselines, allowing agents to act with faulty communication while successfully exploiting shared information.

Gonçalo Querido, Alberto Sardinha, Francisco S. Melo

This work proposes a novel model-free Reinforcement Learning (RL) agent that is able to learn how to complete an unknown task having access to only a part of the input observation. We take inspiration from the concepts of visual attention and active perception that are characteristic of humans and tried to apply them to our agent, creating a hard attention mechanism. In this mechanism, the model decides first which region of the input image it should look at, and only after that it has access to the pixels of that region. Current RL agents do not follow this principle and we have not seen these mechanisms applied to the same purpose as this work. In our architecture, we adapt an existing model called recurrent attention model (RAM) and combine it with the proximal policy optimization (PPO) algorithm. We investigate whether a model with these characteristics is capable of achieving similar performance to state-of-the-art model-free RL agents that access the full input observation. This analysis is made in two Atari games, Pong and SpaceInvaders, which have a discrete action space, and in CarRacing, which has a continuous action space. Besides assessing its performance, we also analyze the movement of the attention of our model and compare it with what would be an example of the human behavior. Even with such visual limitation, we show that our model matches the performance of PPO+LSTM in two of the three games tested.

Fábio Vital, Miguel Vasco, Alberto Sardinha et al.

We present Perceive-Represent-Generate (PRG), a novel three-stage framework that maps perceptual information of different modalities (e.g., visual or sound), corresponding to a sequence of instructions, to an adequate sequence of movements to be executed by a robot. In the first stage, we perceive and pre-process the given inputs, isolating individual commands from the complete instruction provided by a human user. In the second stage we encode the individual commands into a multimodal latent space, employing a deep generative model. Finally, in the third stage we convert the multimodal latent values into individual trajectories and combine them into a single dynamic movement primitive, allowing its execution in a robotic platform. We evaluate our pipeline in the context of a novel robotic handwriting task, where the robot receives as input a word through different perceptual modalities (e.g., image, sound), and generates the corresponding motion trajectory to write it, creating coherent and readable handwritten words.

Pedro P. Santos, Alberto Sardinha, Francisco S. Melo

The general-utility Markov decision processes (GUMDPs) framework generalizes the MDPs framework by considering objective functions that depend on the frequency of visitation of state-action pairs induced by a given policy. In this work, we contribute with the first analysis on the impact of the number of trials, i.e., the number of randomly sampled trajectories, in infinite-horizon GUMDPs. We show that, as opposed to standard MDPs, the number of trials plays a key-role in infinite-horizon GUMDPs and the expected performance of a given policy depends, in general, on the number of trials. We consider both discounted and average GUMDPs, where the objective function depends, respectively, on discounted and average frequencies of visitation of state-action pairs. First, we study policy evaluation under discounted GUMDPs, proving lower and upper bounds on the mismatch between the finite and infinite trials formulations for GUMDPs. Second, we address average GUMDPs, studying how different classes of GUMDPs impact the mismatch between the finite and infinite trials formulations. Third, we provide a set of empirical results to support our claims, highlighting how the number of trajectories and the structure of the underlying GUMDP influence policy evaluation.

João G. Ribeiroa, Cassandro Martinhoa, Alberto Sardinhaa et al.

This paper introduces a formal definition of the setting of ad hoc teamwork under partial observability and proposes a first-principled model-based approach which relies only on prior knowledge and partial observations of the environment in order to perform ad hoc teamwork. We make three distinct assumptions that set it apart previous works, namely: i) the state of the environment is always partially observable, ii) the actions of the teammates are always unavailable to the ad hoc agent and iii) the ad hoc agent has no access to a reward signal which could be used to learn the task from scratch. Our results in 70 POMDPs from 11 domains show that our approach is not only effective in assisting unknown teammates in solving unknown tasks but is also robust in scaling to more challenging problems.

Guilherme S. Varela, Alberto Sardinha, Francisco S. Melo

We propose a novel cooperative multi-agent reinforcement learning (MARL) approach for networked agents. In contrast to previous methods that rely on complete state information or joint observations, our agents must learn how to reach shared objectives under partial observability. During training, they collect individual rewards and approximate a team value function through local communication, resulting in cooperative behavior. To describe our problem, we introduce the networked dynamic partially observable Markov game framework, where agents communicate over a switching topology communication network. Our distributed method, DNA-MARL, uses a consensus mechanism for local communication and gradient descent for local computation. DNA-MARL increases the range of the possible applications of networked agents, being well-suited for real world domains that impose privacy and where the messages may not reach their recipients. We evaluate DNA-MARL across benchmark MARL scenarios. Our results highlight the superior performance of DNA-MARL over previous methods.

Pedro P. Santos, Jacopo Silvestrin, Alberto Sardinha et al.

We propose a provably correct Monte Carlo tree search (MCTS) algorithm for solving \textit{risk-aware} Markov decision processes (MDPs) with \textit{entropic risk measure} (ERM) objectives. We provide a \textit{non-asymptotic} analysis of our proposed algorithm, showing that the algorithm: (i) is \textit{correct} in the sense that the empirical ERM obtained at the root node converges to the optimal ERM; and (ii) enjoys \textit{polynomial regret concentration}. Our algorithm successfully exploits the dynamic programming formulations for solving risk-aware MDPs with ERM objectives introduced by previous works in the context of an upper confidence bound-based tree search algorithm. Finally, we provide a set of illustrative experiments comparing our risk-aware MCTS method against relevant baselines.

Pedro P. Santos, Alberto Sardinha, Francisco S. Melo

In this work, we contribute the first approach to solve infinite-horizon discounted general-utility Markov decision processes (GUMDPs) in the single-trial regime, i.e., when the agent's performance is evaluated based on a single trajectory. First, we provide some fundamental results regarding policy optimization in the single-trial regime, investigating which class of policies suffices for optimality, casting our problem as a particular MDP that is equivalent to our original problem, as well as studying the computational hardness of policy optimization in the single-trial regime. Second, we show how we can leverage online planning techniques, in particular a Monte-Carlo tree search algorithm, to solve GUMDPs in the single-trial regime. Third, we provide experimental results showcasing the superior performance of our approach in comparison to relevant baselines.

Fábio Vital, Alberto Sardinha, Francisco S. Melo

Conversational repair is a mechanism used to detect and resolve miscommunication and misinformation problems when two or more agents interact. One particular and underexplored form of repair in emergent communication is the implicit repair mechanism, where the interlocutor purposely conveys the desired information in such a way as to prevent misinformation from any other interlocutor. This work explores how redundancy can modify the emergent communication protocol to continue conveying the necessary information to complete the underlying task, even with additional external environmental pressures such as noise. We focus on extending the signaling game, called the Lewis Game, by adding noise in the communication channel and inputs received by the agents. Our analysis shows that agents add redundancy to the transmitted messages as an outcome to prevent the negative impact of noise on the task success. Additionally, we observe that the emerging communication protocol's generalization capabilities remain equivalent to architectures employed in simpler games that are entirely deterministic. Additionally, our method is the only one suitable for producing robust communication protocols that can handle cases with and without noise while maintaining increased generalization performance levels.

Guilherme S. Varela, Alberto Sardinha, Francisco S. Melo

We investigate the problem of distributed training under partial observability, whereby cooperative multi-agent reinforcement learning agents (MARL) maximize the expected cumulative joint reward. We propose distributed value decomposition networks (DVDN) that generate a joint Q-function that factorizes into agent-wise Q-functions. Whereas the original value decomposition networks rely on centralized training, our approach is suitable for domains where centralized training is not possible and agents must learn by interacting with the physical environment in a decentralized manner while communicating with their peers. DVDN overcomes the need for centralized training by locally estimating the shared objective. We contribute with two innovative algorithms, DVDN and DVDN (GT), for the heterogeneous and homogeneous agents settings respectively. Empirically, both algorithms approximate the performance of value decomposition networks, in spite of the information loss during communication, as demonstrated in ten MARL tasks in three standard environments.

Bernardo Esteves, Miguel Vasco, Francisco S. Melo

We contribute NeuralSolver, a novel recurrent solver that can efficiently and consistently extrapolate, i.e., learn algorithms from smaller problems (in terms of observation size) and execute those algorithms in large problems. Contrary to previous recurrent solvers, NeuralSolver can be naturally applied in both same-size problems, where the input and output sizes are the same, and in different-size problems, where the size of the input and output differ. To allow for this versatility, we design NeuralSolver with three main components: a recurrent module, that iteratively processes input information at different scales, a processing module, responsible for aggregating the previously processed information, and a curriculum-based training scheme, that improves the extrapolation performance of the method. To evaluate our method we introduce a set of novel different-size tasks and we show that NeuralSolver consistently outperforms the prior state-of-the-art recurrent solvers in extrapolating to larger problems, considering smaller training problems and requiring less parameters than other approaches.

Petra Poklukar, Miguel Vasco, Hang Yin et al.

Learning representations of multimodal data that are both informative and robust to missing modalities at test time remains a challenging problem due to the inherent heterogeneity of data obtained from different channels. To address it, we present a novel Geometric Multimodal Contrastive (GMC) representation learning method consisting of two main components: i) a two-level architecture consisting of modality-specific base encoders, allowing to process an arbitrary number of modalities to an intermediate representation of fixed dimensionality, and a shared projection head, mapping the intermediate representations to a latent representation space; ii) a multimodal contrastive loss function that encourages the geometric alignment of the learned representations. We experimentally demonstrate that GMC representations are semantically rich and achieve state-of-the-art performance with missing modality information on three different learning problems including prediction and reinforcement learning tasks.

João G. Ribeiro, Cassandro Martinho, Alberto Sardinha et al.

In this paper, we present a novel Bayesian online prediction algorithm for the problem setting of ad hoc teamwork under partial observability (ATPO), which enables on-the-fly collaboration with unknown teammates performing an unknown task without needing a pre-coordination protocol. Unlike previous works that assume a fully observable state of the environment, ATPO accommodates partial observability, using the agent's observations to identify which task is being performed by the teammates. Our approach assumes neither that the teammate's actions are visible nor an environment reward signal. We evaluate ATPO in three domains -- two modified versions of the Pursuit domain with partial observability and the overcooked domain. Our results show that ATPO is effective and robust in identifying the teammate's task from a large library of possible tasks, efficient at solving it in near-optimal time, and scalable in adapting to increasingly larger problem sizes.

Pedro P. Santos, Diogo S. Carvalho, Alberto Sardinha et al.

We study the interplay between the data distribution and Q-learning-based algorithms with function approximation. We provide a unified theoretical and empirical analysis as to how different properties of the data distribution influence the performance of Q-learning-based algorithms. We connect different lines of research, as well as validate and extend previous results. We start by reviewing theoretical bounds on the performance of approximate dynamic programming algorithms. We then introduce a novel four-state MDP specifically tailored to highlight the impact of the data distribution in the performance of Q-learning-based algorithms with function approximation, both online and offline. Finally, we experimentally assess the impact of the data distribution properties on the performance of two offline Q-learning-based algorithms under different environments. According to our results: (i) high entropy data distributions are well-suited for learning in an offline manner; and (ii) a certain degree of data diversity (data coverage) and data quality (closeness to optimal policy) are jointly desirable for offline learning.

Miguel Vasco, Hang Yin, Francisco S. Melo et al.

This work addresses the problem of sensing the world: how to learn a multimodal representation of a reinforcement learning agent's environment that allows the execution of tasks under incomplete perceptual conditions. To address such problem, we argue for hierarchy in the design of representation models and contribute with a novel multimodal representation model, MUSE. The proposed model learns hierarchical representations: low-level modality-specific representations, encoded from raw observation data, and a high-level multimodal representation, encoding joint-modality information to allow robust state estimation. We employ MUSE as the sensory representation model of deep reinforcement learning agents provided with multimodal observations in Atari games. We perform a comparative study over different designs of reinforcement learning agents, showing that MUSE allows agents to perform tasks under incomplete perceptual experience with minimal performance loss. Finally, we evaluate the performance of MUSE in literature-standard multimodal scenarios with higher number and more complex modalities, showing that it outperforms state-of-the-art multimodal variational autoencoders in single and cross-modality generation.

Ricardo Quinteiro, Francisco S. Melo, Pedro A. Santos

This paper addresses the problem of optimal control using search trees. We start by considering multi-armed bandit problems with continuous action spaces and propose LD-HOO, a limited depth variant of the hierarchical optimistic optimization (HOO) algorithm. We provide a regret analysis for LD-HOO and show that, asymptotically, our algorithm exhibits the same cumulative regret as the original HOO while being faster and more memory efficient. We then propose a Monte Carlo tree search algorithm based on LD-HOO for optimal control problems and illustrate the resulting approach's application in several optimal control problems.

Guilherme S. Varela, Pedro P. Santos, Alberto Sardinha et al.

This article proposes a methodology for the development of adaptive traffic signal controllers using reinforcement learning. Our methodology addresses the lack of standardization in the literature that renders the comparison of approaches in different works meaningless, due to differences in metrics, environments, and even experimental design and methodology. The proposed methodology thus comprises all the steps necessary to develop, deploy and evaluate an adaptive traffic signal controller -- from simulation setup to problem formulation and experimental design. We illustrate the proposed methodology in two simple scenarios, highlighting how its different steps address limitations found in the current literature.

Miguel Vasco, Francisco S. Melo, Ana Paiva

Humans are able to create rich representations of their external reality. Their internal representations allow for cross-modality inference, where available perceptions can induce the perceptual experience of missing input modalities. In this paper, we contribute the Multimodal Hierarchical Variational Auto-encoder (MHVAE), a hierarchical multimodal generative model for representation learning. Inspired by human cognitive models, the MHVAE is able to learn modality-specific distributions, of an arbitrary number of modalities, and a joint-modality distribution, responsible for cross-modality inference. We formally derive the model's evidence lower bound and propose a novel methodology to approximate the joint-modality posterior based on modality-specific representation dropout. We evaluate the MHVAE on standard multimodal datasets. Our model performs on par with other state-of-the-art generative models regarding joint-modality reconstruction from arbitrary input modalities and cross-modality inference.

Manuel Lopes, Francisco Melo

In this paper we propose the first machine teaching algorithm for multiple inverse reinforcement learners. Specifically, our contributions are: (i) we formally introduce the problem of teaching a sequential task to a heterogeneous group of learners; (ii) we identify conditions under which it is possible to conduct such teaching using the same demonstration for all learners; and (iii) we propose and evaluate a simple algorithm that computes a demonstration(s) ensuring that all agents in a heterogeneous class learn a task description that is compatible with the target task. Our analysis shows that, contrary to other teaching problems, teaching a heterogeneous class with a single demonstration may not be possible as the differences between agents increase. We also showcase the advantages of our proposed machine teaching approach against several possible alternatives.

Rui Silva, Miguel Vasco, Francisco S. Melo et al.

In this work we explore the use of latent representations obtained from multiple input sensory modalities (such as images or sounds) in allowing an agent to learn and exploit policies over different subsets of input modalities. We propose a three-stage architecture that allows a reinforcement learning agent trained over a given sensory modality, to execute its task on a different sensory modality-for example, learning a visual policy over image inputs, and then execute such policy when only sound inputs are available. We show that the generalized policies achieve better out-of-the-box performance when compared to different baselines. Moreover, we show this holds in different OpenAI gym and video game environments, even when using different multimodal generative models and reinforcement learning algorithms.

João Ribeiro, Francisco S. Melo, João Dias

In this paper we investigate two hypothesis regarding the use of deep reinforcement learning in multiple tasks. The first hypothesis is driven by the question of whether a deep reinforcement learning algorithm, trained on two similar tasks, is able to outperform two single-task, individually trained algorithms, by more efficiently learning a new, similar task, that none of the three algorithms has encountered before. The second hypothesis is driven by the question of whether the same multi-task deep RL algorithm, trained on two similar tasks and augmented with elastic weight consolidation (EWC), is able to retain similar performance on the new task, as a similar algorithm without EWC, whilst being able to overcome catastrophic forgetting in the two previous tasks. We show that a multi-task Asynchronous Advantage Actor-Critic (GA3C) algorithm, trained on Space Invaders and Demon Attack, is in fact able to outperform two single-tasks GA3C versions, trained individually for each single-task, when evaluated on a new, third task, namely, Phoenix. We also show that, when training two trained multi-task GA3C algorithms on the third task, if one is augmented with EWC, it is not only able to achieve similar performance on the new task, but also capable of overcoming a substantial amount of catastrophic forgetting on the two previous tasks.

Miguel Vasco, Francisco S. Melo, David Martins de Matos et al.

Humans interact in rich and diverse ways with the environment. However, the representation of such behavior by artificial agents is often limited. In this work we present \textit{motion concepts}, a novel multimodal representation of human actions in a household environment. A motion concept encompasses a probabilistic description of the kinematics of the action along with its contextual background, namely the location and the objects held during the performance. Furthermore, we present Online Motion Concept Learning (OMCL), a new algorithm which learns novel motion concepts from action demonstrations and recognizes previously learned motion concepts. The algorithm is evaluated on a virtual-reality household environment with the presence of a human avatar. OMCL outperforms standard motion recognition algorithms on an one-shot recognition task, attesting to its potential for sample-efficient recognition of human actions.

Patricia Alves-Oliveira, Pedro Sequeira, Francisco S. Melo et al.

This work explores a group learning scenario with an autonomous empathic robot. We address two research questions: (1) Can an autonomous robot designed with empathic competencies foster collaborative learning in a group context? (2) Can an empathic robot sustain positive educational outcomes in long-term collaborative learning interactions with groups of students? To answer these questions, we developed an autonomous robot with empathic competencies that is able to interact with a group of students in a learning activity about sustainable development. Two studies were conducted. The first study compares learning outcomes in children across 3 conditions: learning with an empathic robot; learning with a robot without empathic capabilities; and learning without a robot. The results show that the autonomous robot with empathy fosters meaningful discussions about sustainability, which is a learning outcome in sustainability education. The second study features groups of students who interact with the robot in a school classroom for two months. The long-term educational interaction did not seem to provide significant learning gains, although there was a change in game-actions to achieve more sustainability during game-play. This result reflects the need to perform more long-term research in the field of educational robots for group learning.

Francisco Melo, Manuel Lopes

This paper addresses the problem of learning a task from demonstration. We adopt the framework of inverse reinforcement learning, where tasks are represented in the form of a reward function. Our contribution is a novel active learning algorithm that enables the learning agent to query the expert for more informative demonstrations, thus leading to more sample-efficient learning. For this novel algorithm (Generalized Binary Search for Inverse Reinforcement Learning, or GBS-IRL), we provide a theoretical bound on sample complexity and illustrate its applicability on several different tasks. To our knowledge, GBS-IRL is the first active IRL algorithm with provable sample complexity bounds. We also discuss our method in light of other existing methods in the literature and its general applicability in multi-class classification problems. Finally, motivated by recent work on learning from demonstration in robots, we also discuss how different forms of human feedback can be integrated in a transparent manner in our learning framework.