We present a suite of techniques for extending the Partially Observable Monte Carlo Planning algorithm to handle complex multi-agent games. We design the planning algorithm to exploit the inherent structure of the game. When game rules naturally cluster the actions into sets called types, these can be leveraged to extract characteristics and high-level strategies from a sparse corpus of human play. Another key insight is to account for action legality both when extracting policies from game play and when these are used to inform the forward sampling method. We evaluate our algorithm against other baselines and versus ablated versions of itself in the well-known board game Settlers of Catan.

An intelligent agent performs actions in order to achieve its goals. Such actions can either be externally directed, such as opening a door, or internally directed, such as writing data to a memory location or strengthening a synaptic connection. Some internal actions, to which we refer as computations, potentially help the agent choose better actions. Considering that (external) actions and computations might draw upon the same resources, such as time and energy, deciding when to act or compute, as well as what to compute, are detrimental to the performance of an agent. In an environment that provides rewards depending on an agent's behavior, an action's value is typically defined as the sum of expected long-term rewards succeeding the action (itself a complex quantity that depends on what the agent goes on to do after the action in question). However, defining the value of a computation is not as straightforward, as computations are only valuable in a higher order way, through the alteration of actions. This thesis offers a principled way of computing the value of a computation in a planning setting formalized as a Markov decision process. We present two different definitions of computation values: static and dynamic. They address two extreme cases of the computation budget: affording calculation of zero or infinitely many steps in the future. We show that these values have desirable properties, such as temporal consistency and asymptotic convergence. Furthermore, we propose methods for efficiently computing and approximating the static and dynamic computation values. We describe a sense in which the policies that greedily maximize these values can be optimal. We utilize these principles to construct Monte Carlo tree search algorithms that outperform most of the state-of-the-art in terms of finding higher quality actions given the same simulation resources.

RaySearch exhibited its latest advances in oncology software at the 2018 American Society for Radiation Oncology (ASTRO) annual meeting, Oct. 21-24 in San Antonio, Texas, backed by machine learning and user-friendly tools to enable optimal use of clinical resources. The company showed its latest development in machine learning technology and automation in both its RayStation treatment planning system (TPS) and RayCare oncology information system (OIS). With its in-house machine learning department, RaySearch is currently developing machine learning and deep learning prototypes for areas such as treatment plan generation, organ segmentation and target volume estimation. RaySearch is also developing techniques for efficient large-scale data extraction and analysis. The first clinical applications of machine learning are automated treatment planning and automated organ segmentation and will be included in the next RayStation release in December.

In this paper, we present a hierarchical path planning framework called SG-RL (subgoal graphs-reinforcement learning), to plan rational paths for agents maneuvering in continuous and uncertain environments. By "rational", we mean (1) efficient path planning to eliminate first-move lags; (2) collision-free and smooth for agents with kinematic constraints satisfied. SG-RL works in a two-level manner. At the first level, SG-RL uses a geometric path-planning method, i.e., Simple Subgoal Graphs (SSG), to efficiently find optimal abstract paths, also called subgoal sequences. At the second level, SG-RL uses an RL method, i.e., Least-Squares Policy Iteration (LSPI), to learn near-optimal motion-planning policies which can generate kinematically feasible and collision-free trajectories between adjacent subgoals. The first advantage of the proposed method is that SSG can solve the limitations of sparse reward and local minima trap for RL agents; thus, LSPI can be used to generate paths in complex environments. The second advantage is that, when the environment changes slightly (i.e., unexpected obstacles appearing), SG-RL does not need to reconstruct subgoal graphs and replan subgoal sequences using SSG, since LSPI can deal with uncertainties by exploiting its generalization ability to handle changes in environments. Simulation experiments in representative scenarios demonstrate that, compared with existing methods, SG-RL can work well on large-scale maps with relatively low action-switching frequencies and shorter path lengths, and SG-RL can deal with small changes in environments. We further demonstrate that the design of reward functions and the types of training environments are important factors for learning feasible policies.

There is huge demand for robots to work alongside humans in heterogeneous teams. To achieve a high degree of fluidity, robots must be able to (1) recognize their human co-worker's intent, and (2) adapt to this intent accordingly, providing useful aid as a teammate. The literature to date has made great progress in these two areas -- recognition and adaptation -- but largely as separate research activities. In this work, we present a unified approach to these two problems, in which recognition and adaptation occur concurrently and holistically within the same framework. We introduce Pike, an executive for human-robot teams, that allows the robot to continuously and concurrently reason about what a human is doing as execution proceeds, as well as adapt appropriately. The result is a mixed-initiative execution where humans and robots interact fluidly to complete task goals.Key to our approach is our task model: a contingent, temporally-flexible team-plan with explicit choices for both the human and robot. This allows a single set of algorithms to find implicit constraints between sets of choices for the human and robot (as determined via causal link analysis and temporal reasoning), narrowing the possible decisions a rational human would take (hence achieving intent recognition) as well as the possible actions a robot could consistently take (hence achieving adaptation). Pike makes choices based on the preconditions of actions in the plan, temporal constraints, unanticipated disturbances, and choices made previously (by either agent).Innovations of this work include (1) a framework for concurrent intent recognition and adaptation for contingent, temporally-flexible plans, (2) the generalization of causal links for contingent, temporally-flexible plans along with related extraction algorithms, and (3) extensions to a state-of-the-art dynamic execution system to utilize these causal links for decision making.

A common paradigm in classical planning is heuristic forward search. Forward search planners often rely on relatively simple best-first search algorithm, which remains fixed throughout the search process. In this paper, we introduce a novel search framework capable of alternating between several forward search approaches while solving a particular planning problem. Selection of the approach is performed using a trainable stochastic policy. This enables tailoring the search strategy to a particular distribution of planning problems and a selected performance metric, such as the IPC score or running time. We construct a strategy space using five search algorithms and a two-dimensional representation of the planner's state. Strategies are then trained on randomly generated planning problems using policy gradient. Experimental results show that the learner is able to discover domain-specific search strategies, thus improving the planner's performance with respect to the chosen metric.

Many planning techniques have been developed to allow autonomous systems to act and make decisions based on their perceptions of the environment. Among these techniques, HTN ({\it Hierarchical Task Network}) planning is one of the most used in practice. Unlike classical approaches of planning. HTN operates by decomposing task into sub-tasks until each of these sub-tasks can be achieved an action. This hierarchical representation provide a richer representation of planning problems and allows to better guide the plan search and provides more knowledge to the underlying algorithms. In this paper, we propose a new approach of HTN planning in which, as in conventional planning, we instantiate all planning operators before starting the search process. This approach has proven its effectiveness in classical planning and is necessary for the development of effective heuristics and encoding planning problems in other formalism such as CSP or SAT. The instantiation is actually used by most modern planners but has never been applied in an HTN based planning framework. We present in this article a generic instantiation algorithm which implements many simplification techniques to reduce the process complexity inspired from those used in classical planning. Finally we present some results obtained from an experimentation on a range of problems used in the international planning competitions with a modified version of SHOP planner using fully instantiated problems.

In the context of real-time planning, this paper investigates the contributions of two enhancements for selecting actions. First, the agenda-driven planning enhancement ranks relevant atomic goals and solves them incrementally in a best-first manner. Second, the committed jump enhancement commits a sequence of actions to be executed at the following time steps. To assess these two enhancements, we developed a real-time planning algorithm in which action selection can be driven by a goal-agenda, and committed jumps can be done. Experimental results, performed on classical planning problems, show that agenda-planning and committed jumps are clear advantages in the real-time context. Used simultaneously, they enable the planner to be several orders of magnitude faster and solution plans to be shorter.

For social robots to be brought more into widespread use in the fields of companionship, care taking and domestic help, they must be capable of demonstrating social intelligence. In order to be acceptable, they must exhibit socio-communicative skills. Classic approaches to program HRI from observed human-human interactions fails to capture the subtlety of multimodal interactions as well as the key structural differences between robots and humans. The former arises due to a difficulty in quantifying and coding multimodal behaviours, while the latter due to a difference of the degrees of liberty between a robot and a human. However, the notion of reverse engineering from multimodal HRI traces to learn the underlying behavioral blueprint of the robot given multimodal traces seems an option worth exploring. With this spirit, the entire HRI can be seen as a sequence of exchanges of speech acts between the robot and human, each act treated as an action, bearing in mind that the entire sequence is goal-driven. Thus, this entire interaction can be treated as a sequence of actions propelling the interaction from its initial to goal state, also known as a plan in the domain of AI planning. In the same domain, this action sequence that stems from plan execution can be represented as a trace. AI techniques, such as machine learning, can be used to learn behavioral models (also known as symbolic action models in AI), intended to be reusable for AI planning, from the aforementioned multimodal traces. This article reviews recent machine learning techniques for learning planning action models which can be applied to the field of HRI with the intent of rendering robots as socio-communicative.

Abstract--Planning has achieved significant progress in recent years. Among the various approaches to scale up plan synthesis, the use of macro-actions has been widely explored. As a first stage towards the development of a solution to learn online macro-actions, we propose an algorithm to identify useful macroactions based on data mining techniques. The integration in the planning search of these learned macro-actions shows significant improvements over six classical planning benchmarks. Automated planning is an area of Artificial Intelligence that comes up with the challenge of devising systems that can autonomously find a plan to reach a set of goals. In classical planning, a problem is composed of an initial state, a goal specification and a set of actions. From the initial state if the preconditions of an action are satisfied, the action is applicable to the current state.