Log In Sign Up

Hierarchical Reinforcement Learning with Abductive Planning

by   Kazeto Yamamoto, et al.
National Institute of Advanced Industrial Science and Technology
The University of Tokyo

One of the key challenges in applying reinforcement learning to real-life problems is that the amount of train-and-error required to learn a good policy increases drastically as the task becomes complex. One potential solution to this problem is to combine reinforcement learning with automated symbol planning and utilize prior knowledge on the domain. However, existing methods have limitations in their applicability and expressiveness. In this paper we propose a hierarchical reinforcement learning method based on abductive symbolic planning. The planner can deal with user-defined evaluation functions and is not based on the Herbrand theorem. Therefore it can utilize prior knowledge of the rewards and can work in a domain where the state space is unknown. We demonstrate empirically that our architecture significantly improves learning efficiency with respect to the amount of training examples on the evaluation domain, in which the state space is unknown and there exist multiple goals.


Efficient reinforcement learning control for continuum robots based on Inexplicit Prior Knowledge

Compared to rigid robots that are often studied in reinforcement learnin...

Neural Network iLQR: A New Reinforcement Learning Architecture

As a notable machine learning paradigm, the research efforts in the cont...

Spatially and Seamlessly Hierarchical Reinforcement Learning for State Space and Policy space in Autonomous Driving

Despite advances in hierarchical reinforcement learning, its application...

Artificial Neural Networks that Learn to Satisfy Logic Constraints

Logic-based problems such as planning, theorem proving, or puzzles, typi...

What Should I Do Now? Marrying Reinforcement Learning and Symbolic Planning

Long-term planning poses a major difficulty to many reinforcement learni...

Automated Synthesis of Steady-State Continuous Processes using Reinforcement Learning

Automated flowsheet synthesis is an important field in computer-aided pr...

Behavior Priors for Efficient Reinforcement Learning

As we deploy reinforcement learning agents to solve increasingly challen...

1 Introduction

Reinforcement learning (RL) is a class of machine learning problems in which an autonomous agent learns a policy to achieve a given task through trial and error. Automated planning is an area of artificial intelligence that studies how to make efficient plans for achieving a given task with using predefined knowledge. Reinforcement learning and automated planning are complementary to each other and various methods to combine them have been proposed 

[Partalas et al.2008, Grzes and Kudenko2008, Branavan et al.2012, Konidaris et al.2014, Konidaris et al.2015, Leonetti et al.2016, Andersen and Konidaris2017].

Partalas et al. partalas08 classified those methods into two categories: reinforcement learning to speed up automated planning and reinforcement learning to increase domain knowledge for automated planing. In this paper, we focus on the former, and, more specifically, address the problem of hierarchical reinforcement learning with symbol-based planning. We are motivated by the fact that the lack of efficiency in the learning process is an essential problem which has hampered application of reinforcement learning to complicated domains in the real world.

We argue that there are problems with the automated planners employed by existing hierarchical reinforcement learning frameworks if they are to be used in real-life applications. First, most of the widely used planners and inference models are based on the Herbrand theorem and thus need a set of constants as input to generate a Herbrand universe. Therefore, the meaning representation used in planning by such planners must satisfy the condition that, for each term of the predicates, there exists a finite sequence of constant terms. This condition restricts the available meaning representations and often hampers the application of symbolic planners to domains modeled by a partially observable Markov decision process (POMDP). Second, to the best of our knowledge, existing symbolic planners cannot deal with pre-defined knowledge about rewards and hence cannot evaluate the expectations of rewards in planning. Real-life problems often have several goals with different priorities. For instance, the AI controller of a plant may need to consider multiple goals in planning (e.g. safety operation vs profit maximization). Third, classical planners (e.g. STRIPS) can only deal with rules that define actions. Those planners cannot utilize other types of knowledge, such as relationships of a subordinate concept to a main concept. In order to address the above problems, we propose a new symbolic planner that employs ILP-formulated abduction as its symbolic planner. Abduction is a form of inference that is used to find the best explanations to a given observation. The development of efficient inference techniques for abduction in recent years warrants the application of abduction with large knowledge bases to real-life problems. We show that our model can overcome the above issues through experiments on a Minecraft-like evaluation task.

This paper consists of six sections. In Section 2, we introduce the formalism of reinforcement learning and automated reasoning in abduction, and review some previous work on symbolic planning-based hierarchical reinforcement learning. Section 3 describes our abduction-based hierarchical reinforcement learning framework. Section 4 describes our evaluation domain. In Section 5, we report the results of experiments. The final section concludes the paper.

2 Background

This section reviews related work on hierarchical reinforcement learning with symbolic planning and abduction.

2.1 Reinforcement Learning

Reinforcement Learning is a subfield of machine learning that studies how to build an autonomous agent that can learn a good behavior policy through interactions with a given environment. A problem of RL can be formalized as a 4-tuple , where is a set of propositional states, is a set of available actions,

is a function which defines the probability that taking action

in state will result in a transition to state , and defines the reward received when such a transition is made. The problem is called a Markov Decision Process (MDP) if the states are fully observable; otherwise it is called a Partially Observable Markov Decision Process (POMDP).

The learning efficiency of RL decreases as the state space in the target domain becomes larger. This is a major problem in applying RL to large real-life problems. Although various approaches to solve this problem have been proposed, we focus on methods that utilize symbolic automated planners to improve the learning efficiency.

In automated planning, prior knowledge is used to produce plans which would lead the world from its current state to its goal state. Specially, Symbolic Automated Planning methods deal with symbolic rules and generate symbolic plans.

Grounds & Kudenko grounds07 proposed PLANQ to improve the efficiency of RL in large-scale problems. In PLANQ, a STRIPS planner defines the abstract (high-level) behavior and a RL component learns low-level behavior. PLANQ contains multiple Q-learning agents for each high-level action. Each Q-learning agent learns the behavior to achieve the abstract action corresponding to itself. The authors have shown that a PLANQ-learner learns a good policy efficiently through experiments on the domain where an agent moves in a grid world.

Grzes & Kudenko grzes08 proposed a reward shaping method, in which a potential function considers high-level plans generated by a STRIPS planner. In their method, a RL component (i.e. the low-level planner) will receive intrinsic rewards when the agent follows the high-level plan, where the amount of the intrinsic rewards is bigger as the agent’s state corresponds to the later state in the high-level plan. They have shown that their approach helps the RL component to learn a good policy efficiently through experiments on a navigation maze problem.

DARLING [Leonetti et al.2016] is a model that can utilize a symbolic automated planner to constrain the behavior of the agent to reasonable choices. DARLING employs Answer Set Programming as an automated planner in order to make their approach scalable enough to be applied to real-life problems.

2.2 Abduction

Abduction (or abductive reasoning) is a form of inference to find the best explanation to a given observation. More formally, logical abduction is defined as follows:


Background knowledge , and observations , where and are sets of first-order logical formulas.


A hypothesis (explanation) such that , , where is a set of first-order logical formulas.

Typically, there exist several hypotheses that explain . We call each of them a candidate hypothesis. The goal of abduction is to find the best hypothesis among candidate hypotheses according to a specific evaluation measure. Formally, we find , where is a function , which is called the evaluation function. The best hypothesis is called the solution hypothesis. In the literature, several kinds of evaluation functions have been proposed [Hobbs et al.1993, Raghavan and Mooney2010, Inoue et al.2012, Gordon2016].

Although abduction on first-order logic or similarly expressive formal systems is computationally expensive, inference techniques developed in recent years have improved its computational efficiency [Blythe et al.2011, Inoue and Inui2011, Inoue and Inui2012, Yamamoto et al.2015, Inoue and Gordon2017]. Specially, Inoue et al. inoue11,inoue12a proposed a method (called ILP-formulated Abduction

) to formulate the process of finding the solution hypothesis as a problem of Integer Linear Programming (ILP) and showed that their method significantly improves the computational efficiency of abduction. In addition, since ILP-formulated abduction is based on the directed acyclic graph representation for abduction 

[Charniak and Shimony1990] and thus generates a set of candidate hypotheses in the manner similar to graph generation, it does not need the grounding process. This is a strong advantage in comparison to other inference models based on the Herbrand theorem (e.g. Answer Set Programming, Markov Logic Networks [Richardson and Domingos2006]). More specifically, a set of candidate hypotheses is expressed as a directed graph in which each node corresponds to a logical atom, where each candidate hypothesis corresponds to a subset of nodes in the directed graph. In the process to enumerate candidate hypotheses, ILP-formulated abduction constructs the directed graph by applying two kinds of operations to the observation: backward chaining and unification. Backward chaining is an operation that applies a rule backward (i.e. consider that the presupposition may be true if the consequence is true) and adds atoms in the presupposition to the graph. Unification is an operation that unifies two atoms having the same predicate and makes the assumption that each term of an atom is equal to the corresponding term of the other atom. See Inoue and Inui inoue11 for details.

Abduction has been applied to various real-life problems such as discourse understanding [Inoue et al.2012, Ovchinnikova et al.2013, Sugiura et al.2013, Gordon2016], question answering [And et al.2001, Sasaki2003] and automated planning [Shanahan2000, do Lago Pereira and de Barros2004]. Many planning tasks can be formulated as problems of abductive reasoning by giving an observation consisting of the initial state and the goal state. Abduction will find the solution hypothesis explaining why the goal state has been achieved by starting from the initial state. Then the solution hypothesis can be interpreted a plan from the initial state to the goal state.

Figure 1: An example of a solution hypothesis generated by ILP-formulated abduction. Capitalized terms (e.g. and ) are logical constants and the others (e.g. and ) are logical variables. Atoms in a square are conjunctive (i.e. ) and atoms in gray squares are observations. An equality between terms represents a relation between time points corresponding to the terms. Each solid, directed edge represents an operation of backward-chaining in which the tail atoms are hypothesized from the head atoms. Each dotted, undirected edge represents a unification. Each label on a unification edge such as is an equality between arguments led by the unification.

Figure 1 shows an example of a solution hypothesis by ILP-formulated abduction in automated planning. A solution hypothesis is expressed as a directed acyclic graph and thus we can obtain richer information about the inference than one of other inference frameworks. From this graph, we can obtain the plan to get an apple, namely go a grocery and buy an apple.

3 Proposed Architecture

This section describes abduction-based hierarchical reinforcement learning.

Figure 2: The basic structure of our architecture.

Figure 2 shows the basic structure of our architecture. Using a predefined knowledge base, the abduction-based symbolic planner generates plans at the abstract level. The planner based on a reinforcement learning model interprets the plans made by the symbolic planner as a sequence of subgoals (options) and plans a specific action on the next step. This structure is similar to existing hierarchical reinforcement learning methods based on symbolic planners, such as PLANQ. Following previous work, we call the abstraction level on which the symbolic planner works high-level and the abstraction level on which the planner based on the reinforcement learning model works low-level. We use the term the high-level planner to refer to the planner based on abduction, and the term the low-level planner the planner based on the reinforcement learning model.

Here we describe the algorithm for choosing an action. First, the system converts the current state and the goal state into an observation in first-order logic for abduction. Next, the system performs abduction for this observation and then makes a high-level plan to achieve the goal state. We use a modified version of the evaluation function in Weighted Abduction [Hobbs et al.1993] in order to obtain a good plan. We describe the details of this evaluation function in Section 3.1. Finally, the system decides the next action by considering the nearest subgoal in the high-level plan. Following hierarchical-DQN [Kulkarni et al.2016], the system gives intrinsic rewards to the low-level component when the subgoal is completed, and thus the low-level component will learn the behavior to achieve subgoals by considering the intrinsic rewards.

One can use an arbitrary method to make a high-level plan from a solution hypothesis in our architecture. In this paper, assuming that the graph structure corresponds to the time order, we make a high-level plan from actions sorted by distance from the goal state. For example, actions in the solution hypothesis shown in Figure 1 may be get-apple, buy-apple, have-money and go-grocery. Sorting them by distance from the goal state, we can obtain a high-level plan { go-grocery, buy-apple, get-apple }. The action have-money is excluded from the high-level plan because it has been already satisfied by the current state.

Using ILP-formulated abduction as the high level planner has several benefits. First, since ILP-formulated abduction does not need a set of constants as input, our architecture can deal with a domain in which the size of state space is unpredictable. In other words, there is no need to consider whether the state space made from the current meaning representation is a closed set or not. When using other logical inference models, it is often hard to find which meaning representation is appropriate for the target domain. This difficulty can be sidestepped by ILP-formulated abduction. Second, this advantage gives our architecture another benefit, namely the ability to make plans of an arbitrary length. This is an advantage over existing logical inference models (e.g. Answer Set Programming). Third, an advantage over classical planners is the ability to use types of knowledge other than action definitions. For instance, in STRIPS, one cannot define rules of relations between objects (e.g. ). Finally, ILP-formulated abduction provides directed graphs as the solution hypothesis. Compared with other logical inference models which just return sets of logical symbols as outputs, abduction can provide more interpretable outputs.

3.1 Evaluation Function

In this section, we describe the evaluation function used in the abductive planner of our architecture.

In general abduction, evaluation functions are used to evaluate the plausibility of each hypothesis as the explanation for the observation. For instance, the evaluation function of probabilistic abduction (e.g. Etcetera Abduction [Gordon2016]

) is the posterior probability

, where is a hypothesis and is the observation.

However, what we expect abduction to find in this work is not the most probable one, but the most promising one as a high-level plan. In other words, our evaluation function needs to consider not only the possibility of a hypothesis but also the reward that the agent will receive by completing the plan made from the hypothesis.

Therefore, we add a new term of expected reward to the standard evaluation function:


where is some evaluation function in an existing abduction model, such as Weighted Abduction and Etcetera Abduction. is the task-specific function that evaluates the amount of reward on completing a plan in hypothesis . More specifically, in this paper, we use an evaluation function based on Weighted Abduction:


where is the cost function in Weighted Abduction and is the amount of reward of completing a plan in hypothesis .

Although we employ Weighted Abduction as a base model due to the availability of an efficient reasoning engine, a different abduction model could be used. For instance, using a probabilistic abduction model, one can define an evaluation function to evaluate the exact expectation of reward, namely .

4 Evaluation Domain

This section describes the domain we used for evaluating our abduction-based RL method.

In this paper, we use a domain of grid-based virtual world based on Minecraft. Each grid cell is either of land or lava and can contain materials or utilities. The player can move around the world, pick up materials and craft objects with utilities. Each episode will end when the player arrives at the goal position, when the player walks into a lava-grid, or when the player has executed 100 actions.

In order to examine the effectiveness of our approach empirically, we set up the problem so that it has types of complexity that tend to exist in real-world problems: partial observability, multiple goals, delayed reward and multitask.

Figure 3: An example of tasks in our evaluation domain.
Partial Observability

The player at the initial state does not have any knowledge of the environment. More specifically, he does not know the size of the grid world, where he is, what items there are, or the positions of materials and utilities in the grid world. He can detect the existence of an object in the world when he gets close to it. For example, the gray area around the player in Figure 3 shows the range of his sensing. Therefore, he knows nothing about the outside of this area at the initial state.

Multiple Goals and Delayed Reward

The player receives a reward only when he arrives at the goal position. The amount of the reward depends on what he can craft on arriving at the goal. The reward will be high if he can craft an object made from many materials. For instance, the reward given to a player who has enough materials to cook rabbit-stew is much higher than that for a player who has only collected rabbit.


In this domain, the layout of the grid world is randomized on every episode. Specifically, the player’s starting position, the goal position, the arrangement of lava, the width of the grid-world, the variation of materials and their positions vary randomly. The range of the width of the grid-world is . Each grid world contains kinds of materials and is always surrounded by lava.

It should be noted that, since the variation of materials in the world may change, it is possible that the player cannot craft the optimal object in some episodes. In other words, the optimal goals of different episodes are different. Therefore, the player needs to judge which goal is the most appropriate in each episode. For example, the player will receive the highest reward when he can cook rabbit-stew, which is made from rabbit, bowl, mushroom, potato, carrot and some fuel to use a furnace. Therefore, the player cannot cook it if any of its materials does not exist in the world.

These characteristics make it difficult to apply existing reinforcement learning models to this evaluation domain. The state space in this domain is unpredictable and thus existing first-order symbolic planners based on Herbrand’s theorem, such as Answer Set Programming and STRIPS, are not straightforwardly applicable to this domain.

Let us discuss in more detail the difficulty of applying Herbrand’s theorem-based planners to this domain. Since those planners need a set of constants to make a Herbrand universe, one must define predicates so that one can enumerate all possible arguments in advance. However, most of the objects in this domain (e.g., grid cells, materials and time points) are not enumerable; that is, one cannot define closed sets of arguments corresponding to those objects in advance. Therefore one cannot avoid giving a huge set of constants to deal with all possible cases or abstracting predicates so that its argument set is known in advance. The former may be computationally intractable and the latter may be too time-consuming and difficult for human.

Following the conditions in General Game Playing [Genesereth et al.2005], we had conducted evaluation on the following presuppositions. First, the player can use prior knowledge of the dynamics of the target domain. That includes knowledge of crafting rules and the amount of reward for each object. Second, the player cannot use the knowledge of task-specific strategies for the target domain. In other words, we do not add any knowledge of how to move for getting higher rewards.

5 Experiments

We evaluated our approach in the domain described in Section 4.

We compared the following three models. First, NO-PLANNER is a RL model without a high-level planner. Second, FIXED-GOAL is the model in which the high-level planner always makes plans so that the player achieve the most ideal goal (i.e. cooking rabbit-stew). We consider this model to correspond to existing symbolic planner-based hierarchical RL models, in which the high-level planners cannot deal with prior knowledge of the rewards. Finally, ABDUCTIVE is our proposed model, in which the high-level planner is based on abduction.

We employed Proximal Policy Optimization algorithm (PPO) [Schulman et al.2017]111We used the library implemented by OpenAI, available at as the low-level component for each model. PPO is a state-of-the-art policy gradient RL algorithm, and is relatively easy to implement.

In order to perform abduction based on our evaluation function proposed in Section 3.1, we implemented a modified version of Phillip222It is available at, a state-of-the-art engine for ILP-formulated abductive reasoning, and used it for the high-level planner of ABDUCTIVE. In order to improve the time efficiency of planning, we cached the inference results for each observation and reused them whenever possible.

Figure 4: Examples of rules that we used.

We manually constructed the prior knowledge of the evaluation domain for the high-level planner. The knowledge base contains 31 predicates and 125 rules. As stated in Section 4, these rules consist of only the ones for the dynamics of the domain, such as crafting rules and properties of objects. We describe examples of the rules in Figure 4.

We used roughly three types of actions as elements in high-level plans, namely finding a certain object, picking up a certain material and going to a certain place. Each of them takes one argument (e.g. get-rabbit) and thus we actually used 20 actions in a high-level plan. For instance, a high-level planner in this experiment may generate high-level plans like { find-coal, get-coal, get-rabbit, go-furnace, go-goal }.

As stated previously, the content of each task is generated randomly. Since we use the number of episodes as the random number seed for task generation, the set of tasks in each experiment is the same. We tested the model performance every 10 episodes in each experiment.

Experimental Result

Figure 5 illustrates the curves of the cumulative rewards averaged over 5 independent trials for each model. Each plot is averaged over a sliding window of 500 episodes.

Figure 5: The performance of three models.

As we can see, ABDUCTIVE obtained much more rewards than other models and learned more efficiently with respect to the number of training examples. Learning efficiency is important when RL is applied to real-life problems. In such domains, the time required for trial and error can be prohibitively long because of the computational cost of a simulator or necessity of manual operations.

Figure 6: An example of solution hypotheses generated by the abductive planner in the evaluation domain.

Figure 6 is an example of solution hypotheses made by the abductive planner in the evaluation domain. Our architecture may convert this into a subgoal sequence — pick up rabbit, go to furnace and go to goal. As we can see, since it is described as a graph how the planner inferred the plan, our system can improve interpretability of the content of the inference. From this proof graph, we can see that our planner can make plans of an arbitrary length and can use types of knowledge other than action definitions.

One limitation of our architecture is the large variance in CPU time required per time step. Most of the action selections can be made in a few milliseconds, but when the high-level planner needs to perform abductive planning, the selection may take a few seconds. For this issue, there are several directions of future work to reduce the frequency of performing abduction. One is to improve the algorithm for finding reusable cached results. Our current implementation uses cached results only when exactly the same observation is given. The computational cost of high-level planning could be significantly reduced if the high-level planner can reuse a cache for similar observations as well.

6 Conclusion

We proposed an architecture of abduction-based hierarchical reinforcement learning and demonstrated that it improves the efficiency of reinforcement learning in a complex domain.

Our ILP-formulated abduction-based symbolic planner is not based on the Herbrand theorem and thus can work in domains where the state space is unknown. Moreover, since it can deal with various evaluation functions including user-defined ones, we can easily allow an abductive planner to utilize prior knowledge about rewards.

In future work, we plan to employ machine learning methods for abduction. In recent years, some methods for machine learning of abduction have been proposed [Yamamoto et al.2013, Inoue et al.2012]. Although we manually made prior knowledge for the experiments in this work, we could apply these methods to our architecture. Specifically, if we could divide the errors into the high-level component’s errors and the low-level component’s errors, we can update the weights of symbolic rules used in the abductive planner discriminatively when the high-level planner fails.


  • [And et al.2001] Debra Burhans And, Debra T. Burhans, and Stuart C. Shapiro. Abduction and question answering. In Proceedings of the IJCAI-01 Workshop on Abductive Reasoning. AAAI, 2001.
  • [Andersen and Konidaris2017] Garrett Andersen and George Konidaris. Active exploration for learning symbolic representations. In Neural Information Processing Systems 30, pages 5009–5019, 2017.
  • [Blythe et al.2011] James Blythe, Jerry R. Hobbs, Pedro Domingos, Rohit J. Kate, and Raymond J. Mooney. Implementing weighted abduction in markov logic. In Proceedings of the Ninth International Conference on Computational Semantics, IWCS ’11, pages 55–64, Stroudsburg, PA, USA, 2011. Association for Computational Linguistics.
  • [Branavan et al.2012] S. R. K. Branavan, Nate Kushman, Tao Lei, and Regina Barzilay. Learning high-level planning from text. In Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics: Long Papers - Volume 1, ACL ’12, pages 126–135, Stroudsburg, PA, USA, 2012. Association for Computational Linguistics.
  • [Charniak and Shimony1990] Eugene Charniak and Solomon E. Shimony. Probabilistic semantics for cost based abduction. Proceedings of the Eighth National Conference on Artificial Intelligence (AAAI-90), pages 106–111, 1990.
  • [do Lago Pereira and de Barros2004] Silvio do Lago Pereira and Leliane Nunes de Barros. Planning with abduction: A logical framework to explore extensions to classical planning. In Ana L. C. Bazzan and Sofiane Labidi, editors, Advances in Artificial Intelligence, pages 62–72, Berlin, Heidelberg, 2004. Springer Berlin Heidelberg.
  • [Genesereth et al.2005] Michael Genesereth, Nathaniel Love, and Barney Pell. General game playing: Overview of the AAAI competition. 26:62–72, 06 2005.
  • [Gordon2016] Andrew S. Gordon. Commonsense interpretation of triangle behavior. In Proceedings of the Thirtieth AAAI Conference on Artificial Intelligence, AAAI’16, pages 3719–3725. AAAI Press, 2016.
  • [Grounds and Kudenko2007] Matthew Jon Grounds and Daniel Kudenko. Combining reinforcement learning with symbolic planning. In Adaptive Agents and Multi-Agents Systems, volume 4865 of Lecture Notes in Computer Science, pages 75–86. Springer, 2007.
  • [Grzes and Kudenko2008] Marek Grzes and Daniel Kudenko. Plan-based reward shaping for reinforcement learning. In Fourth IEEE International Conference on Intelligent Systems, 2008.
  • [Hobbs et al.1993] Jerry R. Hobbs, Mark E. Stickel, Douglas E. Appelt, and Paul Martin. Interpretation as abduction. Artificial Intelligence, 63(1):69–142, 1993.
  • [Inoue and Gordon2017] Naoya Inoue and Andrew S. Gordon. A scalable weighted Max-SAT implementation of propositional Etcetera Abduction. In Proceedings of the 30th International Conference of the Florida AI Society (FLAIRS-30), Marco Island, Florida, May 2017. AAAI Press.
  • [Inoue and Inui2011] Naoya Inoue and Kentaro Inui. ILP-based reasoning for weighted abduction. In Proceedings of AAAI Workshop on Plan, Activity and Intent Recognition, pages 25–32, August 2011.
  • [Inoue and Inui2012] Naoya Inoue and Kentaro Inui. Large-scale cost-based abduction in full-fledged first-order predicate logic with cutting plane inference. In Proceedings of the 13th European Conference on Logics in Artificial Intelligence, JELIA’12, pages 281–293, Berlin, Heidelberg, 2012. Springer-Verlag.
  • [Inoue et al.2012] Naoya Inoue, Ekaterina Ovchinnikova, Kentaro Inui, and Jerry Hobbs. Coreference resolution with ILP-based weighted abduction. In 24th International Conference on Computational Linguistics - Proceedings of COLING 2012: Technical Papers, pages 1291–1308, December 2012.
  • [Konidaris et al.2014] George Konidaris, Leslie Pack Kaelbling, and Tomas Lozano-Perez. Constructing symbolic representations for high-level planning. In Proceedings of the Twenty-Eighth AAAI Conference on Artificial Intelligence, AAAI’14, pages 1932–1940. AAAI Press, 2014.
  • [Konidaris et al.2015] George Konidaris, Leslie Pack Kaelbling, and Tomas Lozano-Perez. Symbol acquisition for probabilistic high-level planning. In Proceedings of the 24th International Conference on Artificial Intelligence, IJCAI’15, pages 3619–3627. AAAI Press, 2015.
  • [Kulkarni et al.2016] Tejas D Kulkarni, Karthik Narasimhan, Ardavan Saeedi, and Josh Tenenbaum. Hierarchical deep reinforcement learning: Integrating temporal abstraction and intrinsic motivation. In D. D. Lee, M. Sugiyama, U. V. Luxburg, I. Guyon, and R. Garnett, editors, Advances in Neural Information Processing Systems 29, pages 3675–3683. Curran Associates, Inc., 2016.
  • [Leonetti et al.2016] Matteo Leonetti, Luca Iocchi, and Peter Stone. A synthesis of automated planning and reinforcement learning for efficient, robust decision-making. Artificial Intelligence, 241:103–130, September 2016.
  • [Ovchinnikova et al.2013] Ekaterina Ovchinnikova, Andrew S. Gordon, and Jerry R. Hobbs. Abduction for Discourse Interpretation: A Probabilistic Framework. In Proceedings of the Joint Symposium on Semantic Processing, pages 42–50, November 2013.
  • [Partalas et al.2008] Ioannis Partalas, Dimitrios Vrakas, and Ioannis Vlahavas. Reinforcement learning and automated planning: A survey. In Dimitrios Vrakas and Ioannis Vlahavas, editors, Advanced Problem Solving Techniques. IGI Global, 2008.
  • [Raghavan and Mooney2010] Sindhu Raghavan and Raymond Mooney.

    Bayesian abductive logic programs.

    In Proceedings of the 6th AAAI Conference on Statistical Relational Artificial Intelligence, AAAIWS’10-06, pages 82–87. AAAI Press, 2010.
  • [Richardson and Domingos2006] Matthew Richardson and Pedro Domingos. Markov logic networks. Mach. Learn., 62(1-2):107–136, February 2006.
  • [Sasaki2003] Yutaka Sasaki. Question answering as abduction : A feasibility study at NTCIR QAC1. IEICE transactions on information and systems, 86(9):1669–1676, sep 2003.
  • [Schulman et al.2017] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. CoRR, abs/1707.06347, 2017.
  • [Shanahan2000] Murray Shanahan. An abductive event calculus planner. Journal of Logic Programming, 44:207–239, 2000.
  • [Sugiura et al.2013] Jun Sugiura, Naoya Inoue, and Kentaro Inui. Recognizing implicit discourse relations through abductive reasoning with large-scale lexical knowledge. In

    Proceedings of the 1st Workshop on Natural Language Processing and Automated Reasoning (NLPAR)

    , number 1044 in CEUR Workshop Proceedings, pages 76–87, A Corunna, Spain, 2013.
  • [Yamamoto et al.2013] Kazeto Yamamoto, Naoya Inoue, Yotaro Watanabe, Naoaki Okazaki, and Kentaro Inui. Discriminative learning of first-order Weighted Abduction from partial discourse explanations. In Alexander Gelbukh, editor, Computational Linguistics and Intelligent Text Processing, pages 545–558, Berlin, Heidelberg, 2013. Springer Berlin Heidelberg.
  • [Yamamoto et al.2015] Kazeto Yamamoto, Naoya Inoue, Kentaro Inui, Yuki Arase, and Jun’ichi Tsujii.

    Boosting the efficiency of first-order abductive reasoning using pre-estimated relatedness between predicates.

    5:114–120, april 2015.