# executionguided_neural_program_synthesis__97635c48.pdf

Published as a conference paper at ICLR 2019

EXECUTION-GUIDED NEURAL PROGRAM SYNTHESIS

Xinyun Chen UC Berkeley Chang Liu Citadel Securities Dawn Song UC Berkeley

Neural program synthesis from input-output examples has attracted an increasing interest from both the machine learning and the programming language community. Most existing neural program synthesis approaches employ an encoder-decoder architecture, which uses an encoder to compute the embedding of the given inputoutput examples, as well as a decoder to generate the program from the embedding following a given syntax. Although such approaches achieve a reasonable performance on simple tasks such as Flash Fill, on more complex tasks such as Karel, the state-of-the-art approach can only achieve an accuracy of around 77%. We observe that the main drawback of existing approaches is that the semantic information is greatly under-utilized. In this work, we propose two simple yet principled techniques to better leverage the semantic information, which are execution-guided synthesis and synthesizer ensemble. These techniques are general enough to be combined with any existing encoder-decoder-style neural program synthesizer. Applying our techniques to the Karel dataset, we can boost the accuracy from around 77% to more than 90%.

1 INTRODUCTION

Program synthesis is a traditional challenging problem. Such a problem typically takes a speciﬁcation as the input, and the goal is to generate a program within a target domain-speciﬁc language (DSL). One of the most interesting forms of the speciﬁcations is input-output examples, and there have been several applications, such as Flash Fill (Gulwani, 2011; Gulwani et al., 2012).

Recently, there is an increasing interest of applying neural network approaches to tackle the program synthesis problem. For example, Devlin et al. have demonstrated that using an encoder-decoderstyle neural network, their neural program synthesis algorithm called Robust Fill can outperform the performance of the traditional non-neural program synthesis approach by a large margin on the Flash Fill task (Devlin et al., 2017b).

Despite their promising performance, we identify several inefﬁciencies of such encoder-decoder-style neural program synthesis approaches. In particular, such a neural network considers program synthesis as a sequence generation problem; although some recent work take the syntactical information into consideration during program generation (Bunel et al., 2018; Rabinovich et al., 2017; Yin & Neubig, 2017; Parisotto et al., 2017; Xu et al., 2017), the semantic information, which is typically well-deﬁned in the target DSL, is not effectively leveraged by existing work.

In light of this observation, in this work, we develop simple yet principled techniques that can be combined with any existing encoder-decoder-style neural program synthesizers. The main novel technique is called execution-guided synthesis. The basic idea is to view the program execution as a sequence of manipulations to transform each input state into the corresponding output state. In such a view, executing a partial program can result in intermediate states; thus, synthesizing the rest of the program can be conditioned on these intermediate states, so that the synthesizer can take the state changes into account in the followup program generation process. Therefore, we can leverage this idea to combine with any existing encoder-decoder-style neural synthesizer, and we observe that it can signiﬁcantly improve the performance of the underlying synthesizers.

In addition, we also propose a simple yet effective technique called synthesizer ensemble, which leverages the semantic information to ensemble multiple neural program synthesizers. In particular, for the input-output program synthesis problem, we can easily verify if a synthesized program satisﬁes the input-output speciﬁcation, which allows us to remove invalid predictions from the ensemble

Published as a conference paper at ICLR 2019

during inference time. Albeit its simplicity, to the best of our knowledge, we are not aware of any previous neural program synthesis work applying this approach. We observe that this technique further boosts the performance substantially.

We evaluate our techniques on the Karel task (Bunel et al., 2018; Devlin et al., 2017a), the largest publicly available benchmark for input-output program synthesis, on which the most performant model in the past can achieve only an accuracy of around 77% (Bunel et al., 2018). We observe that our proposed techniques can gain better performance than the previous state-of-the-art results. In particular, by combining both of our techniques, we can achieve an accuracy of 92%, which is around 15 percentage points better than the state-of-the-art results. This shows that our approach is effective in boosting the performance of algorithms for neural program synthesis from input-output examples.

2 NEURAL PROGRAM SYNTHESIS FROM INPUT-OUPUT EXAMPLES

In this section, we ﬁrst introduce the input-output program synthesis problem and existing encoderdecoder-style neural program synthesis approaches, then present an overview of our approaches.

2.1 PROBLEM DEFINITION

We follow the literature (Devlin et al., 2017b; Bunel et al., 2018; Chen et al., 2018) to formally deﬁne the input-output program synthesis problem below.

Problem Deﬁnition 1 (Program emulation) Let L be the space of all valid programs in the domainspeciﬁc language (DSL). Given a set of input-output pairs {(Ik, Ok)}K k=1 (or {IOK} in short), where there exists a program P L, such that P(Ik) = Ok, k {1, ..., K}. Our goal is to compute the output Otest for a new test input Itest, so that Otest = P(Itest).

Although the problem deﬁnition only requires to compute the output for a test input, a typical method is to synthesize a program P L such that P is consistent with all input-output pairs {IOK}, and then use P to compute the output. In this case, we say program P emulates the program P corresponding to {IOK}.

In particular, in this work, we are mainly interested in the following formulation of the problem.

Problem Deﬁnition 2 (Program synthesis) Let L be the space of all valid programs in the domainspeciﬁc language (DSL). Given a training dataset of {IOK}i for i = 1, ..., N, where N is the size of the training data, compute a synthesizer Γ, so that given a test input-output example set {IOK}test, the synthesizer Γ({IOK}test) = P produces a program P, which emulates the program corresponding to {IOK}test.

2.2 ENCODER-DECODER-STYLE NEURAL PROGRAM SYNTHESIS APPROACHES

There have been many approaches proposed for different neural program synthesis tasks, and most of them follow an encoder-decoder-style neural network architecture (Bunel et al., 2018; Devlin et al., 2017b; Parisotto et al., 2017). Figure 1 shows a general neural network architecture for input-output program synthesis. First, an encoder converts input-output examples {IOK} into an embedding. For example, Robust Fill (Devlin et al., 2017b) deals with the string transformation tasks, thus it uses LSTMs as the encoder. For the Karel task, both inputs and outputs are 2D grids (see Figure 2); therefore, Bunel et al. (2018) employ a CNN as the encoder.

Once the IO embeddings are computed, these approaches employ an LSTM decoder to generate the programs conditioned on the embeddings. For program synthesis, one unique property is that the generated program should satisfy the synthax of L. Therefore, a commonly used reﬁnement is to ﬁlter syntactically invalid program preﬁxes during generation (Parisotto et al., 2017; Devlin et al., 2017b; Bunel et al., 2018).

In the above approaches, only the syntax information is leveraged; the semantics of L is not utilized. In particular, standard supervised training procedure could suffer from program aliasing: for the same input-output examples, there are multiple semantically equivalent programs, but all except the one provided in the training data will be penalized as wrong programs. To mitigate this issue, Bunel et

Published as a conference paper at ICLR 2019

Figure 1: A neural network architecture for input-output program synthesis (e.g., (Bunel et al., 2018)). At each timestep t, the decoder LSTM generates a program token gt conditioned on both the input-output pairs {IOK} and the previous program token gt 1. Each IO pair is fed into the LSTM individually, and a max-pooling operation is performed over the hidden states {hk t }K k=1 of the last layer of LSTM for all IO pairs. The resulted vector is fed into a softmax layer to obtain a prediction probability distribution over all the possible program tokens in the vocabulary. More details can be found in Appendix C.

Figure 2: An example of the execution of partial programs to reach the target state in the Karel domain. The blue dot denotes the marker put by the Karel robot.

al. propose to train with reinforcement learning, so that it rewards all semantically correct programs once they are fully generated (Bunel et al., 2018). In our work, we demonstrate that we can leverage the semantic information in an effective way that provides a better performance.

2.3 AN OVERVIEW OF OUR APPROACHES

In this work, we propose two general and principled techniques that can improve the performance over existing work, which are execution-guided synthesis (Section 3) and synthesizer ensemble (Section 4). The main idea of our techniques is to better leverage the semantics of the language L during synthesis. Meanwhile, our techniques are compatible with any existing encoder-decoder-style neural program synthesis architecture. We will describe these techniques in detail in the following sections.

3 EXECUTION-GUIDED SYNTHESIS

Existing approaches generate the entire program only based on the input-output examples before execution. However, this is an inefﬁcient use of the semantics of L. For example, when a program consists of a sequence of statements, we can view the output to be a result by continuously executing each statement in the sequence to convert the input state into a sequence of intermediate states. Figure 2 illustrates such an example. From this perspective, instead of generating the whole program at once, we can generate one statement at a time based on the intermediate/output state pairs.

However, most interesting programs are not just sequential. In this work, we explore this idea using a general control-ﬂow framework. In particular, given any language L, we extend it with three classical types of control-ﬂow: sequential, branching, and looping. The extended language is call Lext. Then, we develop our above idea based on Lext, called execution-guided synthesis. In the following, to make our discussion concise, we ﬁrst formalize Lext (Section 3.1), and then present the idea of execution-guided synthesis (Section 3.2).

Published as a conference paper at ICLR 2019

S-Stmt S, s s

S, s , s S-Seq B1, s B 1, s B1; B2, s B 1; B2, s

S-Seq-Bot ; B2, s B2, s S-If C, s b i = 1 b is true 2 b is false

if C then B1 else B2 ﬁ, s Bi, s

S-While while C do B end, s if C then B; while C do B end else ﬁ, s

Table 2: Semantic rules B, s B , s for Lext.

3.1 THE FORMAL SPECIFICATION OF THE EXTENDED LANGUAGE Lext

P := B; B := | S | B; B | if C then B else B ﬁ | while C do B end S, C L

Table 1: Syntax of Lext.

In this work, we assume some additional control-ﬂow syntax on top of L. We deﬁne the extended language Lext in Table 1. In particular, we assume that a code block B can be composed by a sequence of statements S L or sub code blocks, and each code block can also be an if-statement or a while-statement. We use C to indicate a condition expression in L, and to indicate the termination of a program execution.

The semantics of Lext is speciﬁed in Table 2. These rules are largely standard following the convention in programming language literature. In particular, each rule s name starts with Sindicating that they are semantics rules; the sufﬁxes indicate the constructors each rule speciﬁes, (e.g., Stmt for statements, Seq for sequences, etc.). The judgment B, s B , s indicates a small-step execution of program B over state s to result in a new program B and a new state s . The judgments S, s s and C, s b capture the big-step execution in L that statement S evaluates to a new state s from s, and condition C evaluates to a boolean value b from s. Following the semantics of Lext, we can formally deﬁne a program execution.

Deﬁnition 1 (Program execution) Given a program P Lext and an input I, the execution is a sequence s0...s T , such that (1) s0 = I; (2) B0 = P; (3) Bi, si Bi+1, si+1 for i = 0, ..., T 1; and (4) BT = . The output of the program is O = s T .

3.2 EXECUTION-GUIDED SYNTHESIS ALGORITHM

In Deﬁnition 1, we can observe that the initial and ﬁnal states are simply two special states provided as the input-output examples of the synthesis problem. Thus, a synthesizer Γ for input-output pairs should also be able to take any state-pairs as inputs. Our execution-guided synthesis algorithm takes advantage of this fact to improve the performance of the synthesizer. In the following, we discuss three cases from the easiest to the hardest to present our approach.

Sequential programs. We now consider the simplest case, where the program is in the form of S1; ...; ST , to illustrate the basic idea of execution-guided synthesis algorithm. We present the algorithm in Algorithm 1. Assuming the input-output examples are {IOK}, we can treat them as K state-pairs {(sk i , sk o)}K k=1, where sk i = Ik, sk o = Ok. The Exec algorithm takes the synthesizer Γ and the input-output pairs {IOK} as its input. It also takes an additional input , which is the ending token to be synthesized. For the top-level program, will be the token. Later we will see that when synthesizing the sub-program for Ifand While-blocks, different ending tokens will be used.

The synthesized program is initially empty (line 3). Then the algorithm iteratively generates one statement S at a time (line 4-10 and 17), and appends it to the end of P (line 16). Importantly, if S is not an if-statement or a while-statement, for which we handle separately, the algorithm executes the newly generated statement S to transit sk i into sk new (line 11-14). Therefore, in the subsequent iteration, the synthesizer can start from the new states sk new after executing the partial program P

Published as a conference paper at ICLR 2019

Algorithm 1 Execution-guided synthesis (sequential case)

1: function EXEC(Γ, {(sk i , sk o)}K k=1, ) 2: // The main algorithm is called using Exec (Γ, {IOK}, ) 3: P 4: S Γ({(sk i , sk o)}K k=1) 5: while S = do 6: if S = if-token then // If-statement synthesis 7: S, {(sk i , sk o)}K k=1 Exec If(Γ, {(sk i , sk o)}K k=1) 8: else 9: if S = while-token then // While-statement synthesis 10: S, {(sk i , sk o)}K k=1 Exec While(Γ, {(sk i , sk o)}K k=1) 11: else // Execution of S 12: S, sk i , sk new for k = 1, ..., K 13: sk i sk new for k = 1, ..., K 14: end if 15: end if 16: P P; S 17: S Γ({(sk i , sk o)}K k=1) 18: end while 19: return P 20: end function

Algorithm 2 Execution-guided synthesis (if-statement)

1: function EXECIF(Γ, I) 2: C Γ(I) 3: It {(si, so) I| C, si true} 4: If {(si, so) I| C, si false} 5: Bt Exec(Γ, It, else-token) 6: Bf Exec(Γ, If, ﬁ-token) 7: I t {(snew, so)|(si, so) It Bt, si snew} 8: I f {(snew, so)|(si, so) If Bf, si snew} 9: I I t I f 10: S if C then Bt else Bf ﬁ 11: return S, I 12: end function

generated so far. In doing so, the synthesizer can see all intermediate states to better adjust the followup synthesis strategies to improve the overall synthesis performance.

Branching programs. Dealing with if-statements is slightly more complicated than sequential programs, since in an if-statement, not all statements will be executed on all inputs. Following the execution in Algorithm 1 naively, we have to use Γ to synthesize the entire if-statement before being able to execute the partially generated program to derive intermediate states.

Therefore, in Algorithm 2, we extend the above idea to handle if-statements. When the next predicted token is an if-token, our execution-guided synthesizer ﬁrst predicts the condition of the if-statement C (line 2). Then, we evaluate C over all state-pairs. Based on the evaluation results, we can divide the IO pairs into two sets It and If (line 3-4), so that all states in the former meet the branching condition to go to the true branch, and all states in the latter go to the false branch. Therefore, in the followup synthesis, we do not need to consider If (or It) when synthesizing the true branch (or the false branch) (line 5-6). Note that in line 5-6, synthesizing both true-branch and false-branch employ execution-guided synthesis algorithm to leverage intermediate states, and different ending tokens (i.e., else and ﬁ) are supplied respectively. Once we have done the synthesis of both branches, we can execute the generated branches to get the new states I (line 7-9), and return the newly generated if-statement and updated states to the caller of this algorithm.

Published as a conference paper at ICLR 2019

In Algorithm 2, we use B, s s to indicate a big-step execution of code block B over state s to get s . In particular, this means that B, s B1, s1 ... , s .

Looping programs. The remaining problem is to handle while-statements. Due to the rule S-While (see Table 2), a while-statement while C do B end (1)

is equivalent to if C then (B; while C do B end) else ﬁ (2)

Therefore, we can employ a procedure similar to Algorithm 2 once a while-token is predicted. However, there are two differences. First, in (2), the false-branch is empty, thus we do not need to deal with the false-branch. Second, although the true-branch is B; while C do B end, once we have generated B, we do not need to generate the rest of the true-branch, since both C and B have been generated. The detailed algorithm can be found in Appendix B.

Remarks. The ﬁnal algorithm is called by Exec (Γ, {IOK}, ). Note that our execution-guided synthesis algorithm can be applied to any neural synthesizer Γ, and we can train the synthesizer Γ using any supervised or reinforcement learning algorithms that have been proposed before (Devlin et al., 2017b; Bunel et al., 2018). In our evaluation, we demonstrate that our execution-guided synthesis technique helps boost the performance of both supervised and reinforcement learning algorithms proposed in existing work (Bunel et al., 2018).

4 SYNTHESIZER ENSEMBLE

In our experiments, we observe that when we use different random initializations of the synthesizer during training, even if the synthesizer architectures are the same, they will be effective on different subsets of the dataset, although the overall prediction accuracy is similar to each other. Thus, a natural idea is to train multiple synthesizers, and ensemble them to build a more performant synthesizer.

Different from other deep learning tasks, for program synthesis task, without knowing the ground truth, we can already ﬁlter out those wrong predictions that cannot satisfy the input-output speciﬁcation. Thus, we ensemble multiple synthesizers as follows: we run all synthesizers to obtain multiple programs, and select from programs that are consistent with all input-output examples. This provides us with a better chance to select the correct ﬁnal prediction that generalizes to held-out IO pairs.

The main subtlety of such an approach is to deal with the case when multiple generated programs satisfy the input-output examples. In this work, we consider several alternatives as follows:

Majority vote. We can choose the most frequently predicted program as the ﬁnal prediction.

Shortest. Following the Occam s razor principle, we can choose the shortest program as the ﬁnal prediction.

5 EVALUATION

In this section, we demonstrate the effectiveness of our approaches on the Karel dataset (Pattis, 1981; Bunel et al., 2018). We ﬁrst introduce the task, discuss the experimental details, and present the results.

5.1 THE KAREL TASK

Karel is an educational programming language proposed in the 1980s (Pattis, 1981). Using this language, we can control a robot to move inside a 2D grid world and modify the world state, and our goal is to synthesize a program given a small number of input and output grids as the speciﬁcation. Such tasks have been used in Stanford CS introductory courses (CS106A, 2018) and the Hour of Code (Ho C, 2018), and have been studied recently in several neural program synthesis works (Devlin et al., 2017a; Bunel et al., 2018; Shin et al., 2018). Figure 2 shows an example in the Karel domain. We provide the grammar speciﬁcation and the state representation in Appendix A. In particular, the

Published as a conference paper at ICLR 2019

Training Ensemble Generalization Exact Match From

- 71.91% 39.94% Bunel et al. (2018) S 78.80% 46.68% This work MV 78.80% 47.08% This work

- 77.12% 32.17% Bunel et al. (2018) S 84.84% 46.04% This work MV 84.16% 46.36% This work

- 85.08% 40.88%

S 91.60% 45.84% MV 91.52% 45.36%

- 86.04% 39.40% S 91.68% 46.36% MV 92.00% 45.64%

Table 3: Accuracy on the Karel test set. In the Training column, we use MLE and Exec to indicate the training approaches proposed in (Bunel et al., 2018) and this work, SL and RL to indicate supervised learning and reinforcement learning respectively. In the Ensemble column, dash indicates that no ensemble is used, S and MV indicate the shortest and majority vote principles respectively. For the single model accuracy, we report the results of the model with the best generalization accuracy. We include 15 models in each ensemble.

Karel DSL includes control ﬂow constructs such as conditionals and loops, which is more challenging than problems well-studied before, such as Flash Fill (Gulwani, 2011; Devlin et al., 2017b).

Our evaluation follows the setup in (Bunel et al., 2018). We train and evaluate our approaches on their dataset, which is built by randomly sampling programs from the DSL. For each program, 5 IO pairs serve as the speciﬁcation, and the sixth one is the held-out test sample. In total, there are 1,116,854 programs for training, 2,500 in the validation set, and 2,500 in the test set. We evaluate the following two metrics, which are the same as in (Bunel et al., 2018):

Exact Match. The predicted program is an exact match if it is the same as the ground truth.

Generalization. The predicted program is a generalization if it satisﬁes the input-output examples in both the speciﬁcation and the held-out examples.

Training dataset construction for the Exec algorithm. Note that the original Karel dataset only provides the input-output examples and the ground truth programs. To train the synthesizer Γ with our Exec algorithm in the supervised learning setting, we need the supervision on intermediate states as well, which can be obtained easily by executing the ground truth programs following the semantics (Table 2). In particular, for each sample {IOK}, P in the original training set, we construct a sequence of training samples {(sk i 1, Ok)}Ki k=1, Si (i = 1, 2, ..., T), with each sample containing Ki K state pairs and a program Si P. The algorithm to construct this set is largely analogous to the semantics speciﬁcation, and we defer the details to Appendix B.

Training algorithms. Once the training set is constructed, the neural synthesizer Γ can be trained on this new dataset using the same algorithm as the one for training Γ on the original dataset. Therefore, our Exec algorithm can be applied to both supervised learning and reinforcement learning algorithms proposed in (Bunel et al., 2018) for evaluation.

Model details. We employ the same neural network architecture as in (Bunel et al., 2018) to synthesize the programs, which is brieﬂy discussed in Section 2.2. During the inference time, we set the beam size B = 64, and select the one with the highest prediction probability from the remaining programs. More details can be found in Appendix C.

5.2 RESULTS

We present our main results in Table 3. We report the results of ensembling 15 models for our ensemble techniques. For reference, we include MLE and RL results in (Bunel et al., 2018), which were the state-of-the-art on the Karel task for the exact match and generalization metrics respectively.

Published as a conference paper at ICLR 2019

Figure 3: Results of the ensemble model trained with Exec + RL approach. Left: generalization accuracy. Right: exact match accuracy. The corresponding ﬁgures using models trained with Exec approach can be found in Appendix D.2.

We ﬁrst apply our ensemble techniques to these approaches, and observe that the performance could be signiﬁcantly boosted by around 7%.

We next observe that our execution-guided synthesis alone can signiﬁcantly improve the generalization accuracy over all approaches from (Bunel et al., 2018), even after we accompany their approaches with our ensemble techniques. In particular, without the ensemble, Exec+SL already improves MLE+RL by 8 points on generalization accuracy; and when ensemble approaches are applied to MLE+RL , this gap is shrunk, but still positive. Similar to (Bunel et al., 2018), we can also train our Exec model using the RL technique, which improves the generalization accuracy by another 1 point, while slightly decreases the exact match accuracy. These results show that utilizing intermediate execution states alone is already an effective approach to boost the performance.

Note that the improvement of Exec on the exact match accuracy is relatively minor, and sometimes negative when applying the ensemble to baseline training algorithms. This is because our Exec algorithm is not designed to optimize for exact match accuracy. In particular, we decouple the full programs in the original training dataset into small pieces, thus our synthesizer Γ is trained with segments of the original training programs instead of the full programs. In doing so, our synthesizer is more capable of generating programs piece-by-piece and thus tends to generate semantically correct programs (i.e., with a better generalization accuracy) rather than the same programs in the training and testing sets (i.e., with a better exact match accuracy). In fact, for real-world applications, the generalization accuracy is more important than the exact match accuracy, because exact match accuracy is more about evaluating how well the synthesizer recovers the language model of the p CFG sampler used to generate the dataset. More discussion can be found in Appendix D.1.

Finally, we apply our ensemble approaches on top of Exec+SL and Exec+RL. We observe that this can further improve the generalization accuracy and exact match accuracy by around 6% on top of the best single model. These results show that our ensemble approaches consistently boost the performance, regardless of the underlying models used for ensembling.

In addition, we investigate the performance of ensembling different number of models. We present the results of ensembling Exec + RL models in Figure 3, and defer the results of ensembling Exec models to Appendix D.2. We observe that using the shortest principle is generally more effective than using the majority vote principle, especially when fewer number of models are included in the ensemble. However, when there are more models, majority vote may achieve a better generalization accuracy than the shortest principle. This is reasonable, since when there are too few models, there might not be enough effective models to form the majority.

Interestingly, we observe that Exec+RL+Ensemble does not signiﬁcantly improve the performance over Exec+SL+Ensemble. This may be due to that the improvement from ensemble hides the improvement from RL. More discussion of our evaluation results can be found in Appendix D.

To summarize, we make the following key observations:

1. Our execution-guided synthesis technique can effectively improve previous approaches, which only use the syntactic information, or the ﬁnal program execution outputs.

Published as a conference paper at ICLR 2019

2. Our ensemble approaches can effectively improve the performance regardless of the underlying models being used.

3. The different modules of our proposed approaches, i.e., execution-guided synthesis and ensemble techniques, can work independently to improve the performance, and thus they can be applied independently to other tasks as well.

4. By combining all our novel techniques, we improve the state-of-the-art on the Karel task by 14.88 points (generalization) and 7.14 points (exact match).

6 RELATED WORK

Synthesizing a program from input-output examples is an important challenging problem with many applications (Devlin et al., 2017b; Gulwani et al., 2012; Gulwani, 2011; Bunel et al., 2018; Chen et al., 2018; Cai et al., 2017; Li et al., 2017; Reed & De Freitas, 2016; Zaremba et al., 2016; Zaremba & Sutskever, 2015; Fox et al., 2018; Xiao et al., 2018; Ganin et al., 2018). There has been an emerging interest in studying neural networks for program synthesis. A line of work studies training a neural network to directly generate the outputs given the inputs (Devlin et al., 2017b;a; Graves et al., 2014; Joulin & Mikolov, 2015; Kaiser & Sutskever, 2015). In particular, Devlin et al. study the Karel domain (Devlin et al., 2017a). However, as shown in (Devlin et al., 2017a), this approach is incapable of handling the case when the number of input-output examples is small, and is hard to generalize to unseen inputs.

Recent work study using neural networks to generate programs in a domain-speciﬁc language (DSL) from a few input-output examples (Devlin et al., 2017b; Bunel et al., 2018; Parisotto et al., 2017; Polosukhin & Skidanov, 2018; Zohar & Wolf, 2018). Several work synthesize programs for Flash Fill tasks, which are in the string transformation domain (Devlin et al., 2017b; Parisotto et al., 2017). Other work synthesize programs in a LISP-style DSL for array manipulation (Polosukhin & Skidanov, 2018; Zohar & Wolf, 2018). In particular, Zohar & Wolf (2018) also study the idea of encoding the state of the transformed inputs as it is updated during execution. However, these DSLs only include sequential programs, and do not support more complex control ﬂows such as loops and conditionals in our studied Karel problem. Prior works also consider incorporating syntax constraints and information from program execution to facilitate program synthesis (Devlin et al., 2017b; Wang et al., 2018a; Bunel et al., 2018). However, all these works generate the whole program, and use its execution results to guide the synthesis process; in contrast, our work leverages more ﬁne-grained yet generic semantic information that can be gathered during executing programs in most imperative languages. As a result, our approach s performance is signiﬁcantly better than previous work (Bunel et al., 2018).

Previous work also study program synthesis given intermediate states. For example, Sun et al. (2018) propose to synthesize the program from demonstration videos, which can be viewed as sequences of states. In such a problem, all intermediate states can be extracted from the videos. On the contrary, in the input-output program synthesis problem studied in our work, the input to the program synthesizer provides only the initial state and the ﬁnal state. Thus, our synthesizer is required to address the challenge of inferring intermediate states, which is mainly tackled by our execution-guided synthesis algorithm.

In contrast to training a neural network to generate the entire program, a recent line of research studies using a neural network to guide the symbolic program search based on the input-output speciﬁcation, so that the search process prioritizes the operators that have higher domain-speciﬁc scores predicted by the neural networks (Balog et al., 2017; Vijayakumar et al., 2018). Instead of predicting such domain-speciﬁc scores to guide the program search, we directly incorporate the domain knowledge by executing partial programs, and utilize the execution results for program generation of the neural network synthesizer. Meanwhile, recent work propose to leverage the full execution traces in the context of program repair (Wang et al., 2018b). Our work is the ﬁrst to leverage partial execution traces for the program synthesis task, which is a much harder task than program repair.

Published as a conference paper at ICLR 2019

7 CONCLUSION

In this work, we propose two general and principled techniques to better leverage the semantic information for neural program synthesis: (1) execution-guided synthesis; and (2) synthesizer ensemble. On a rich DSL with complex control ﬂows, we achieve a signiﬁcant performance gain over the existing work, which demonstrates that utilizing the semantic information is crucial in boosting the performance of neural program synthesis approaches. We believe that our techniques are also beneﬁcial to other program generation applications, and we consider extending our techniques to handle programming languages with richer semantics as important future work. At the same time, we have observed that utilizing existing reinforcement learning techniques does not provide much performance gain when combined with our approaches. We believe that there is plenty of room left for further improvement, and we are also interested in exploring this problem in the future.

ACKNOWLEDGEMENT

We thank the anonymous reviewers for their valuable comments. This material is in part based upon work supported by the National Science Foundation under Grant No. TWC-1409915, Berkeley Deep Drive, and DARPA D3M under Grant No. FA8750-17-2-0091. Any opinions, ﬁndings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reﬂect the views of the National Science Foundation.

Matej Balog, Alexander L Gaunt, Marc Brockschmidt, Sebastian Nowozin, and Daniel Tarlow. Deepcoder: Learning to write programs. In ICLR, 2017.

Rudy Bunel, Matthew Hausknecht, Jacob Devlin, Rishabh Singh, and Pushmeet Kohli. Leveraging grammar and reinforcement learning for neural program synthesis. In International Conference on Learning Representations, 2018.

Jonathon Cai, Richard Shin, and Dawn Song. Making neural programming architectures generalize via recursion. In ICLR, 2017.

Xinyun Chen, Chang Liu, and Dawn Song. Towards synthesizing complex programs from inputoutput examples. In ICLR, 2018.

CS106A. Stanford cs106a course page. https://see.stanford.edu/Course/CS106A, 2018.

Jacob Devlin, Rudy R Bunel, Rishabh Singh, Matthew Hausknecht, and Pushmeet Kohli. Neural program meta-induction. In Advances in Neural Information Processing Systems, pp. 2077 2085, 2017a.

Jacob Devlin, Jonathan Uesato, Surya Bhupatiraju, Rishabh Singh, Abdel-rahman Mohamed, and Pushmeet Kohli. Robustﬁll: Neural program learning under noisy I/O. In ICML, 2017b.

Roy Fox, Richard Shin, Sanjay Krishnan, Ken Goldberg, Dawn Song, and Ion Stoica. Parametrized hierarchical procedures for neural programming. In ICLR, 2018.

Yaroslav Ganin, Tejas Kulkarni, Igor Babuschkin, SM Eslami, and Oriol Vinyals. Synthesizing programs for images using reinforced adversarial learning. ar Xiv preprint ar Xiv:1804.01118, 2018.

Alex Graves, Greg Wayne, and Ivo Danihelka. Neural turing machines. ar Xiv preprint ar Xiv:1410.5401, 2014.

Sumit Gulwani. Automating string processing in spreadsheets using input-output examples. In ACM SIGPLAN Notices, volume 46, pp. 317 330. ACM, 2011.

Sumit Gulwani, William R Harris, and Rishabh Singh. Spreadsheet data manipulation using examples. Communications of the ACM, 55(8):97 105, 2012.

Ho C. Hour of code. https://codehs.com/hourofcode/, 2018.

Published as a conference paper at ICLR 2019

Armand Joulin and Tomas Mikolov. Inferring algorithmic patterns with stack-augmented recurrent nets. In NIPS, 2015.

Łukasz Kaiser and Ilya Sutskever. Neural gpus learn algorithms. ar Xiv preprint ar Xiv:1511.08228, 2015.

Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015.

Chengtao Li, Daniel Tarlow, Alexander Gaunt, Marc Brockschmidt, and Nate Kushman. Neural program lattices. In ICLR, 2017.

Emilio Parisotto, Abdel-rahman Mohamed, Rishabh Singh, Lihong Li, Dengyong Zhou, and Pushmeet Kohli. Neuro-symbolic program synthesis. In ICLR, 2017.

Richard E Pattis. Karel the robot: a gentle introduction to the art of programming. John Wiley & Sons, Inc., 1981.

Illia Polosukhin and Alexander Skidanov. Neural program search: Solving programming tasks from description and examples. ar Xiv preprint ar Xiv:1802.04335, 2018.

Maxim Rabinovich, Mitchell Stern, and Dan Klein. Abstract syntax networks for code generation and semantic parsing. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), volume 1, pp. 1139 1149, 2017.

Scott Reed and Nando De Freitas. Neural programmer-interpreters. In ICLR, 2016.

Richard Shin, Illia Polosukhin, and Dawn Song. Towards speciﬁcation-directed program repair. 2018.

Shao-Hua Sun, Hyeonwoo Noh, Sriram Somasundaram, and Joseph Lim. Neural program synthesis from diverse demonstration videos. In International Conference on Machine Learning, pp. 4797 4806, 2018.

Ashwin J Vijayakumar, Abhishek Mohta, Oleksandr Polozov, Dhruv Batra, Prateek Jain, and Sumit Gulwani. Neural-guided deductive search for real-time program synthesis from examples. In ICLR, 2018.

Chenglong Wang, Kedar Tatwawadi, Marc Brockschmidt, Po-Sen Huang, Yi Mao, Oleksandr Polozov, and Rishabh Singh. Robust text-to-sql generation with execution-guided decoding. ar Xiv preprint ar Xiv:1807.03100, 2018a.

Ke Wang, Rishabh Singh, and Zhendong Su. Dynamic neural program embedding for program repair. In ICLR, 2018b.

Ronald J Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. In Reinforcement Learning, pp. 5 32. Springer, 1992.

Da Xiao, Jo-Yu Liao, and Xingyuan Yuan. Improving the universality and learnability of neural programmer-interpreters with combinator abstraction. In ICLR, 2018.

Xiaojun Xu, Chang Liu, and Dawn Song. Sqlnet: Generating structured queries from natural language without reinforcement learning. ar Xiv preprint ar Xiv:1711.04436, 2017.

Pengcheng Yin and Graham Neubig. A syntactic neural model for general-purpose code generation. In ACL, 2017.

Wojciech Zaremba and Ilya Sutskever. Reinforcement learning neural turing machines-revised. ar Xiv preprint ar Xiv:1505.00521, 2015.

Wojciech Zaremba, Tomas Mikolov, Armand Joulin, and Rob Fergus. Learning simple algorithms from examples. In Proceedings of The 33rd International Conference on Machine Learning, pp. 421 429, 2016.

Amit Zohar and Lior Wolf. Automatic program synthesis of long programs with a learned garbage collector. In Advances in Neural Information Processing Systems, pp. 2094 2103, 2018.

Published as a conference paper at ICLR 2019

A MORE DESCRIPTIONS OF THE KAREL DOMAIN

Figure 4 presents the grammar speciﬁcation of the Karel DSL.

Each Karel grid world has a maximum size of 18 18, and is represented as a 16 18 18 tensor, where each cell of the grid is represented as a 16-dimensional vector corresponding to the features described in Table 4.

Prog p ::= def run() : s Stmt s ::= while(b) : s | repeat(r) : s | s1 ; s2 | a | if(b) : s | ifelse(b) : s1 else : s2 Cond b ::= front Is Clear() | left Is Clear() | right Is Clear | markers Present() | no Markers Present() | not b Action a ::= move() | turn Right() | turn Left() | pick Marker() | put Marker() Cste r ::= 0 | 1 | ... | 19

Figure 4: Grammar for the Karel task.

Robot facing North

Robot facing East Robot facing South

Robot facing West

Obstacle Grid boundary

1 marker 2 markers 3 markers 4 markers 5 markers 6 markers 7 markers 8 markers 9 markers 10 markers

Table 4: Representation of each cell in the Karel state.

B MORE DETAILS ABOUT THE EXECUTION-GUIDED ALGORITHM

While-statement synthesis algorithm. Algorithm 3 demonstrates the execution-guided algorithm for while-statement synthesis.

Training dataset construction for supervised learning. Consider a program P = S1; ...; ST ; , where each Si is in one of the following forms: (1) Si L; (2) if C then Bt else Bf ﬁ; and (3) while C do B end. For each Si L, we construct a sample of {(sk i 1, Ok)}K k=1, Si directly.

For Si = if C then Bt else Bf ﬁ, we ﬁrst construct a training sample {(sk i 1, Ok)}K k=1, if C then . Afterwards, we split the input-output examples into two subsets It If, where for all (s, O) It, we have C, s true; and C evaluates to false for all (s, O) If on the other hand. Then we obtain two derived samples It, Bt else; Si+1; ...; ST ; and If, Bf ﬁ; Si+1; ...; ST ; , from which we construct training samples respectively using the same approach as discussed above.

In a similar way, we can deal with Si = while C do B end. Finally, we include {(Ok, Ok)}K k=1, in our constructed training set.

Published as a conference paper at ICLR 2019

Algorithm 3 Execution-guided synthesis (while-statement)

1: function EXECWHILE(Γ, I) 2: C Γ(I) 3: It {(si, so) I| C, si true} 4: If {(si, so) I| C, si false} 5: Bt Exec(Γ, It, end-token) 6: I t {(snew, so)|(si, so) It Bt, si snew} 7: I I t If 8: S while C do Bt end 9: return S, I 10: end function

C MODEL DETAILS

C.1 NEURAL NETWORK ARCHITECTURE

Our neural network architecture can be found in Figure 1, which follows the design in Bunel et al. (2018). In particular, the IO Encoder is a convolutional neural network to encode the input and output grids, which outputs a 512-dimensional vector for each input-output pair. The decoder is a 2-layer LSTM with a hidden size of 256. The embedding size of the program tokens is 256.

Each program is represented as a sequence of tokens G = [g1, g2, ..., g L], where each program token gi belongs to a vocabulary Σ. At each timestep t, the decoder LSTM generates a program token gt conditioned on both the input-output pair and the previous program token gt 1, thus the input dimension is 768. Each IO pair is fed into the LSTM individually, and we a max-pooling operation is performed over the hidden states {ht}K k=1 of the last layer of LSTM for all IO pairs. The resulted 256-dimensional vector is fed into a softmax layer to obtain a prediction probability distribution over all the 52 possible program tokens in the vocabulary.

Notice that this neural network architecture can also be applied to other program synthesis problems, with modiﬁcations of the IO encoder architectures for different formats of input-output pairs. For example, in the domain where input-output examples are text strings, such as Flash Fill (Gulwani, 2011), the IO encoders can be recurrent neural networks (RNNs) (Devlin et al., 2017b).

C.2 TRAINING OBJECTIVE FUNCTIONS

To estimate the parameters θ of the neural network, we ﬁrst perform supervised learning to maximize the conditional log-likelihood of the referenced programs (Parisotto et al., 2017; Devlin et al., 2017b; Bunel et al., 2018). In particular, we estimate θ such that

θ = arg max θ

i=1 pθ(πi|{IOk i }K k=1) = arg max θ

i=1 log pθ(πi|{IOk i }K k=1) (3)

Where πi are the ground truth programs provided in the training set.

When training with reinforcement learning, we leverage the policy gradient algorithm REINFORCE (Williams, 1992) to solve the following objective:

θ = arg max θ

G log pθ(G|{IOk i }K k=1)Ri(G) (4)

Where Ri(G) is the reward function to represent the quality of the sampled program G. In our evaluation, we set Ri(G) = 1 if G gives the correct outputs for given inputs, and Ri(G) = 0 otherwise.

Published as a conference paper at ICLR 2019

Figure 5: An example of the predicted program that generalizes to all input-output examples, but is different from the ground truth. Here, we only include 2 out of 5 input-output examples for simplicity. Notice that the predicted program is simpler than the ground truth.

C.3 TRAINING HYPER-PARAMETERS

We use the Adam optimizer (Kingma & Ba, 2015) for both the supervised training and the RL training. The learning rate of supervised training is 10 4, and the learning rate of reinforcement learning is 10 5. We set the batch size to be 128 for supervised training, and 16 for RL training.

D EVALUATION DETAILS

D.1 MORE ANALYSIS OF EVALUATION RESULTS

In our evaluation, we observe that while our Exec approach signiﬁcantly boosts the generalization accuracy, the performance gain of the exact match accuracy is much smaller, and sometimes even negative. We attribute this to the fact that the ground truth program in the Karel benchmark is not always the simplest one satisfying the input-output examples; on the other hand, our approach tends to provide short predictions among all programs consistent with the input-output speciﬁcation. For example, Figure 5 shows a predicted program that is simpler than the ground truth, while also satisﬁes the input-output pairs. Notice that different from the MLE approach in (Bunel et al., 2018), our model is not directly optimized for the exact match accuracy, since the training set not only includes the input-output examples in the original training set, but also the intermediate state pairs, which constitute a larger part of our augmented training set. Meanwhile, in our training set, for state pairs resembling {(sk 1, sk 2)}K k=1 in Figure 5, it is more common for the ground truth program to be a single move() statement than other more complicated ones. Therefore, when training with our approach, the model is less prone to overﬁtting to the sample distribution of the original dataset, and focuses more on the program semantics captured by the intermediate execution.

D.2 MORE DETAILS OF THE ENSEMBLE

For different training approaches of a single model, we train 15 models with different random initializations. To do the ensemble, we ﬁrst sort the 15 models according to the descending order of their generalization accuracies on the validation set, then select the ﬁrst k models to compute the results of the k-model ensemble. When multiple programs satisfy the ensemble criterion, e.g., with the shortest length for the Shortest method, we choose the one from the models with better generalization accuracies on the validation set.

Figure 6 shows the results of the ensemble trained with our Exec approach. Tables 5 and 6 show the numerical results of applying ensemble to our Exec approach and Exec + RL approach.

Published as a conference paper at ICLR 2019

Figure 6: Results of the ensemble model trained with our Exec approach. Left: generalization accuracy. Right: exact match accuracy.

Ensemble 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Exec (S) 40.88% 42.16% 42.64% 43.12% 43.00% 43.36% 44.32% 44.64% 44.72% 45.64% 45.32% 45.60% 45.52% 45.92% 45.84% Exec (MV) 40.88% 41.24% 41.36% 41.52% 41.96% 41.76% 42.64% 43.12% 43.48% 43.56% 43.92% 44.44% 44.80% 45.32% 45.36% Exec + RL (S) 39.40% 42.80% 43.56% 43.84% 44.32% 44.96% 45.16% 45.44% 45.52% 46.36% 46.24% 46.04% 46.28% 46.16% 46.04% Exec + RL (MV) 39.40% 40.76% 41.56% 42.92% 42.84% 43.84% 43.68% 44.36% 44.48% 44.60% 45.24% 45.52% 45.48% 45.64% 45.64%

Table 5: Exact match accuracy of the ensemble.

Ensemble 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Exec (S) 85.08% 86.44% 87.04% 87.52% 87.76% 88.20% 89.04% 89.72% 90.16% 90.28% 90.68% 91.16% 91.44% 91.40% 91.60% Exec (MV) 85.08% 86.08% 86.84% 87.36% 87.72% 87.84% 88.68% 89.36% 89.80% 90.00% 90.44% 91.12% 91.48% 91.40% 91.52% Exec + RL (S) 86.04% 87.20% 88.64% 89.40% 89.64% 90.24% 90.32% 90.44% 90.48% 90.60% 90.96% 91.24% 91.32% 91.64% 91.68% Exec + RL (MV) 86.04% 86.88% 88.24% 89.16% 89.20% 90.04% 90.08% 90.20% 90.56% 90.68% 91.28% 91.48% 91.76% 91.96% 92.00%

Table 6: Generalization accuracy of the ensemble.