Gathering information about the environment state is the main goal in several planning tasks for autonomous agents, such as surveillance, inspection and tracking of objects. Such planning tasks are typically modeled using a Partially Observable Markov Decision Process (POMDP), and in the literature several approaches have emerged to consider information gathering during planning and execution. Similar developments can be seen in the field of active inference, which focuses on active information collection in order to be able to reach a goal. Both fields use POMDPs to model the environment, but the underlying principles for action selection are different. In this paper we create a bridge between both research fields by discussing how they relate to each other and how they can be used for information gathering. Our contribution is a tailored approach to model information gathering tasks directly in the active inference framework. A series of experiments demonstrates that our approach enables agents to gather information about the environment state. As a result, active inference becomes an alternative to common POMDP approaches for information gathering, which opens the door towards more cross cutting research at the intersection of both fields. This is advantageous, because recent advancements in POMDP solvers may be used to accelerate active inference, and the principled active inference framework may be used to model POMDP agents that operate in a neurobiologically plausible fashion.@en

In this paper, the novel Distributed Bayesian (D-Bay) algorithm is presented for solving multi-agent problems within the Continuous Distributed Constraint Optimization Problem (C-DCOP) framework. This framework extends the classical DCOP framework towards utility functions with continuous domains. D-Bay solves a C-DCOP by utilizing Bayesian optimization for the adaptive sampling of variables. We theoretically show that D-Bay converges to the global optimum of the C-DCOP for Lipschitz continuous utility functions. The performance of the algorithm is evaluated empirically based on the sample efficiency. The proposed algorithm is compared to state-of-the-art DCOP and C-DCOP solvers. The algorithm generates better solutions while requiring fewer samples.@en

Robotic systems operating in the real world would benefit from a clear semantic model to understand their interactions with the real world. Such semantics are typically captured in an ontology. Unfortunately, existing world models in robotics focus on its navigation task. They adopt a hierarchical structure decomposing the environment from large spaces into small objects having a position, thereby limiting the robot's interactions as a 'go-to-object' task. To allow a richer understanding of the real world this hierarchical structure should be replaced with an ontology, yet one that does not limit the real-time requirements of the robot when it is queried or updated with new observations extracted from sensors. Such an ontology is presented in this article. For now the ontology also focusses on the navigation aspect of robots, yet it is open to model other aspects of the real world as well. Experiments show that multiple environments are successfully modelled supporting the robot to go from one room to another to search for humans.

@en

A novel partitioning approach for linear switching large-scale systems is presented. We assume that the modes of the switching system are unknown a priori but can be detected. We propose an online partitioning scheme that can partition the system when the mode switches, thus adapting the partition to the mode. Moreover, after the system has been partitioned, we apply a decentralized state-feedback control scheme to stabilize the system. We also apply a dwell time stability scheme to prove that the closed-loop system remains stable even after both the mode and partition changes. The proposed approach is illustrated by means of an automatic generation control problem related to frequency deviation regulation in a large-scale power network.

@en

In the context of building heating systems control in office buildings, the current state-of-the-art applies either a deterministic Model Predictive Control (MPC) controller together with a nonlinear model, or a linearized model with a stochastic MPC controller. Deterministic MPC considers only one realization of the external disturbances, which can lead to a low performance solution if the forecasts of the disturbances are not accurate. Similarly, linear models are simplified representations of the building dynamics and might fail to capture some relevant behavior. In this paper, we improve upon the current literature by combining these two approaches, i.e. we adopt a nonlinear model together with a stochastic MPC controller. We consider a scenario-based MPC (SBMPC), where many realizations of the disturbances are considered, so as to include more possible future trajectories for the external disturbances. The adopted scenario generation method provides statistically significant scenarios, whereas so far in the current literature only approximate methods have been applied. Moreover, we use Modelica to obtain the model description, which allows to have a more accurate and nonlinear model. Lastly, we perform simulations comparing standard MPC vs SBMPC vs an optimal control approach with measurements of the external disturbances, and we show how our proposed scenario-based MPC controller can achieve a better performance compared to standard deterministic MPC.

@en

A single-level rule-based model predictive control (RBMPC) scheme is presented for optimizing the energy management of a grid-connected microgrid composed of local production units, renewable energy sources, local loads, and several types of energy storage systems (ESSs). The single-level controller uses two different models that yield different descriptions of the microgrid and use different sampling times. The model with a smaller sampling time provides a more detailed description of the microgrid, in order to keep track of the fast dynamics, while the model with a higher sampling time provides a less detailed description and is used for making long-term predictions when it is not needed anymore to track the fast dynamics. Moreover, we propose a novel RBMPC method that assigns the value to the binary decision variables in the hybrid microgrid model, e.g., ON or OFF status of the generators and charging or discharging mode of ESSs, through if-then-else rules, which rely on the price of electricity and the local net imbalance. The standard method of applying model predictive control (MPC) to a hybrid model results in a mixed-integer linear programming (MILP) problem. Our proposed rule-based method is able to convert the standard MILP problem into a linear one. We compare our approach through simulations to the MILP approach and show that our method yields almost no loss in performance while providing a significant reduction in the computation time.@en

In this work, the novel algorithm Bayesian Dynamic Programming Optimization Procedure (B-DPOP) is presented to solve multi-agent problems within the Distributed Constraint Optimization Problem framework. The Bayesian optimization framework is used to prove convergence to the global optimum of the B-DPOP algorithm for Lipschitz-continuous objective functions. The proposed algorithm is assessed based on the benchmark problem known as dynamic sensor placement. Results show increased performance over related algorithms in terms of sample-efficiency.

@en

In this paper, Mine Counter-Measures (MCM) operations with multiple cooperative Autonomous Underwater Vehicles (AUVs) are examined within the Distributed Constraint optimization Problem (DCOP) framework. The goal of an MCM-operation is to search for mines and mine-like objects within a predetermined area so that ships can pass the area through a safe transit corridor. Performance metrics, such as the expected time of completion and the level of confidence that all mine-like objects within the area have been detected, are used to quantity the utility of the operation. The AUVs coordinate their individual search segments in a distributed manner in order to maximize the global utility. The segmentation is optimized by the Compression-DPOP (C-DPOP) algorithm, which allows explicit reasoning by the AUVs about their actions based on the performance metrics. After initial segmentation of the mine threat area, subsequent optimizations are triggered by the AUVs based on the variations in sonar performance. The performance of the C-DPOP algorithm is compared to a static segmentation approach and validated using the high-fidelity Unmanned Underwater Vehicle (UUV) simulation environment based on the Gazebo simulator@en

We propose a parametrized Model Predictive Control (MPC) approach for optimal operation of microgrids. The parametrization expresses the control input as a function of the states, variables, and parameters. In this way, it is possible to apply an MPC approach by optimizing only the parameters and not the inputs. Moreover, the value of the binary control variables in the model is assigned according to parametrized heuristic rules, thus obtaining a formulation for the optimization problem that is more scalable compared to standard approaches in the literature. Furthermore, we propose a control scheme based on one single controller that uses two different sampling times and prediction models. By doing so, we can include both fast and slow dynamics of the system at the same level. This control approach is applied to an operational control problem of a microgrid, which includes local loads, local production units, and local energy storage systems and results show the effectiveness of the proposed appro.@en

DCOP (Distributed Constraint optimization Problem) is a framework for representing distributed multi- agent problems. However, it only allows discrete values for the decision variables, which limits its application for real-world problems. In this paper, an extension of DCOP is investigated to handle variables with continuous domains. Additionally, an iterative any-time algorithm Compression-DPOP (C-DPOP) is presented that is based on the Distributed Pseudo-tree Opti- mization Procedure (DPOP). C-DPOP iteratively samples the search space in order to handle problems that are restricted by time and memory limitations. The performance of the algorithm is examined through a mobile sensor coordination problem. The proposed algorithm outperforms DPOP with uniform sampling regarding both resource requirement and performance.

@en