Towards Mixed Optimization for Reinforcement Learning with Program Synthesis

Towards Mixed Optimization for
Reinforcement Learning with Program Synthesis

Surya Bhupatiraju    Kumar Krishna Agrawal    Rishabh Singh

Deep reinforcement learning has led to several recent breakthroughs, though the learned policies are often based on black-box neural networks. This makes them difficult to interpret, and to impose desired specification or constraints during learning. We present an iterative framework, Morl, for improving the learned policies using program synthesis. Concretely, we propose to use synthesis techniques to obtain a symbolic representation of the learned policy, which can be debugged manually or automatically using program repair. After the repair step, we distill the policy corresponding to the repaired program, which is further improved using gradient descent. This process continues until the learned policy satisfies the constraints. We instantiate Morl for the simple CartPole problem and show that the programmatic representation allows for high-level modifications which in turn leads to improved learning of the policies.

Reinforcement Learning, Program Synthesis, Machine Learning, ICML

1 Introduction

There have been many recent successes in using deep reinforcement learning (Drl) to solve challenging problems such as learning to play Go and Atari games (Silver et al., 2016, 2017; Mnih et al., 2015). While the effectiveness of reinforcement learning methods in these domains has been impressive, they have some shortcomings. These learned policies are based on black-box deep neural networks which are difficult to interpret. Additionally, it is challenging to impose and validate certain desirable policy specifications, such as worst-case guarantees or safety constraints. This makes it difficult to debug and improve these policies, therefore hindering their use for safety-critical domains.

There has been some recent work on using program synthesis techniques to interpret learned policies using higher-level programs (Verma et al., 2018) and decision trees (Bastani et al., 2018). The key idea in Pirl (Verma et al., 2018) is to first train a Drl policy using standard methods and then use an imitation learning-like approach to search for a program in a domain-specific language (DSL) that conforms to the behavior traces sampled from the policy. Similarly, Viper (Bastani et al., 2018) uses imitiation learning (a modified form of the Dagger algorithm (Ross et al., 2011)) to extract a decision tree corresponding to the learned policy. The main goal of these works is to extract a symbolic high-level representation of the policy (as a DSL program or a decision tree) which is more interpretable and also amenable for program verification techniques.

We build upon these recent advances to propose an iterative framework for learning interpretable and safe policies. The main steps in the workflow of our framework are as follows. We start with a random initial policy . We use program synthesis techniques similar to Pirl and Viper to learn a symbolic representation of the learned policy as a program . After obtaining a programmatic representation of the policy, we perform program repair (Weimer et al., 2009; Jobstmann et al., 2005) to obtain a repaired program that satisfies some set of constraints. Note that the program repair step can be performed either automatically using a safety specification constraint or it can be performed manually by a human expert that modifies to remove undesirable behaviors (or add desired behaviors). We then use behavioral cloning (Bratko et al., 1995) to obtain the corresponding improved policy , which is then further improved using standard gradient descent to obtain . This process of improving policies from is repeated until achieving desirable performance and safety guarantees. We name this iterative procedure a mixed optimization scheme for reinforcement learning, or Morl.

Figure 1: An overview of the proposed method. We decompose policy learning into alternating between policy optimization and program repair. Starting from a black-box policy , we consider the following steps (1) Synthesis, which generates a program corresponding to the policy . The program is sampled from an underlying Domain Specific Language (DSL) (2) Repair, which corresponds to debugging the program ,allowing us to impose high-level constraints on the learned program. (3) Imitation corresponds to distilling the program back into a reactive representation. (4) Policy Optimization in this case corresponds to gradient-based policy optimization.

As a first step towards a full realization of Morl, we present a simple instantiation of our framework for the CartPole (Barto et al., 1983) problem. We demonstrate the efficacy of our approach to learn near-optimal policies, while enabling the user to better interpret the learned policy. In addition, we argue that the scheme has a natural interpretation and can be readily extended to capture more notions of policy improvement and discuss the potential benefits and obstacles of using such an approach.

This paper makes the following key contributions:

  • We propose a simple framework for iterative policy refinement by performing repair at the level of programmatic representation of learned policies.

  • We instantiate the framework for the CartPole problem and show the effectiveness of performing modifications in the symbolic representation.

2 Mixed Optimization for Reinforcement Learning

Our goal is to improve policy learning by decomposing the usual gradient-based optimization scheme into an iterative two-stage algorithm. In this context, we view improvement as either making the policies (1) safe – to ensure performance under safety, (2) interpretable – allowing some level of introspection into the policy’s decisions, (3) sample efficient, or (4) alignment with priors. While there are other notions of improvement, for the remainder of the paper, we focus on sample efficiency as a notion of policy improvement. We include a discussion of the other approaches as they apply to our framework.

2.1 Problem Definition

Consider the typical Markov decision process (MDP) setup , with a state space , an action space , a reward function , the transition dynamics of the environment , the initial starting state distribution , and the discount factor . The goal will be to find a policy, or function , that achieves the maximum expected reward. Normally, the reward design and specification for a task corresponds to defining the reward function , such that an optimal policy solves the task.

An alternative view of solving the task could be defined as having access to an oracle policy or a fixed number of trajectories from it. In this setting, our goal is learning a policy by imitation learning, which would also equivalently solve the task. In this work, we focus on improving policy learning using imitation learning (Abbeel & Ng, 2004; Ho & Ermon, 2016), though the framework is more general and extends well to reinforcement learning.

We consider a symbolic representation (such as a DSL) that is expressive enough to represent different policies. The synthesis problem can then be defined as learning a program such that , i.e. the learned program produces approximately the same output actions as the actions produced by the policy for all (or a sampled set of) input states .

2.2 Model

In Morl we maintain two representations of a policy:

  • a reactive, black-box policy where we represent the policy as a differentiable function, such as a neural network, allowing us to use gradient-based optimization methods like TRPO (Schulman et al., 2015) or PPO (Schulman et al., 2017).

  • a symbolic program, which represents the policy as an interpretable program. The symbolic program representations are amenable for analysis and transformations using automated program verification and repair techniques, or human inspection.

With these intermediate representations, we alternate between the following; the first step allows us to finetune policies in function space and the second allows us to impose constraints or incorporate human debugging. The procedure (Fig 1) consists of four key steps, as detailed below.

Figure 2: Evaluating the usefulness of maintaining differentiable, and symbolic representations of the policy. Each plot corresponds to finetuning a policy cloned from a program (in this case decision trees) with TRPO (averaged over 5 runs). In this case, Near-Optimal is obtained by manual debugging of the Intermediate policy, which is obtained from Worst policy.

Synthesis: Given a task , we consider a Domain Specific Language , such that there exists some program that is a sufficient representation of the task. In the first step of Morl, we seek to synthesize such a program that is equivalent to the policy . A programmatic representation of the policy allows us to leverage approaches such as program repair and verification to provide guarantees for the underlying policy. For this step, and in the scope of this paper, we assume that we can utilize existing program synthesis methods such as Viper or Pirl, so we do not attempt to perform this step explicitly. We focus on the following steps in the Morl scheme.

Repair: In this step, we modify the synthesized program accordingly to satisfy constraints imposed either on , or on the synthesized program . This step allows us to meaningfully debug the policy, either through human-in-the-loop verification for interpretability, or through automated program repair techniques that involve defining Constraint Satisfaction Problems (CSP) typically solved using SAT/SMT solvers (Singh et al., 2013). For the scope of this paper, we mimic the repair process by manually modifying the initial program to obtain three programs that achieve three different levels of success at the task of interest.

Imitation: Following the program synthesis and repair steps, we distill (Rusu et al., 2015) the program back into a reactive policy using imitation learning. Given that we have access to an oracle , we find that we reliably imitate the program (Ross et al., 2011). Note that it is possible to stop the optimization here. Indeed, we observe that a user may end the procedure of Morl here, if certain performance or safety bounds have been reached, and may skip the last step.

Figure 3: An important step in the algorithm is alternating between symbolic and policy representations. Here we plot the convergence rate of randomly initialized policies to the program behavior. In this work, we used simple behavioral cloning to retrain the policies. We note that more sample efficient algorithms would be able to emulate the behavior from the program more quickly.

Policy Optimization Finally, we finetune the policy using gradient descent. We posit that by optimizing in both program space and over the space of policies in a differentiable space, we are able to better escape local minima while still maintaining an underlying intuition for how the policy is performing from the inspection of the program.

3 Experiments

We evaluate our framework on the CartPole-v0 problem in the OpenAI Gym environment for discrete control (Brockman et al., 2016). We present a first simple instantiation of the framework to showcase its usefulness compared to direct reinforcement learning. In our preliminary evaluation, we evaluate the following research questions:

  • Does program repair lead to faster convergence?

  • Does programmatic representation help humans provide better repair insights?

To this effect, we train an initial policy (Worst) that performs poorly, and then extract the corresponding symbolic representation . For the symbolic representation, we chose Viper’s (Bastani et al., 2018) decision tree representation of the policy. We then modify the symbolic program to get a new program , which performs better than the original program by repairing certain values in the decision tree. This is followed by behavioural cloning to obtain (corresponding to ), which is optimized to obtain .

To simulate the iterative optimization of the framework, we perform two different modifications of the program repair step to obtain (Intermediate) and (Near-optimal) that have different characteristics in terms of repair improvements. For example, the modification to obtain program from is shown in Fig 4, where we manually provide the insight of making the cart shift in the same direction as the pole.

Figure 4: Debugging Worst (red) to Intermediate (green). In one step of debugging the policy, we fix the policy to make the cart shift in the same direction as the pole.

In our experiments, we first find the average performance of each of the levels of policies across 25 runs. The Worst policy gets an average reward of 9.28, the Intermediate policy gets an average reward of 104.0, and the Near-optimal policy gets an average reward of 200.0. When we attempt to distill the programs to continuous policies , we find that each of the resulting levels of policies get 10.64, 66, and 185, respectively, as shown in Figure 3 after 15000 epochs. Lastly, when we take the resulting distilled policies and then finetune these with TRPO, we find that the resulting average rewards are 38.65, 79.03, and and 176.8 after 25 episodes of training with 10 trajectories of length 200. In Figure 2, we run TRPO for a total of 250 episodes to see the limiting behavior.

From our results, we validate our hypothesis that under bad initialization (Worst), TRPO takes an order of magnitude longer to converge to near-optimal policy, when compared to policies initialized after program repair. We believe that providing high-level insights programmatically can help policies discover better or safer behaviors.

4 Related Work

Our framework is inspired from the recent works of Pirl (Verma et al., 2018) and Viper (Bastani et al., 2018) in using program synthesis to learn symbolic interpretable representations of learnt policies, and then using program verification to verify certain properties of the program.

Pirl first trains a Drl policy for a domain and then uses an imitation learning like approach to generate specifications (input-output behaviors) for the synthesis problem. It then uses a Bayesian optimization technique to search for programs in a DSL that conforms to the specification. It iteratively builds up new behaviors by executing the initial policy as an oracle to obtain outputs for inputs that were not originally sampled but are observed in executing the learnt programs. It maintains a family of programs consistent with the specification and chooses the one as output that achieves the maximum reward on the task.

Viper uses a modified form of the Dagger initiation learning algorithm to extract a decision tree corresponding to the learnt policy. It then uses program verification techniques to validate correctness, stability, and robustness properties of the extracted programs (represented as decision trees).

While previous approaches stop at learning a verifiable symbolic representation of policies, our framework aims at iterative improvement of policies. In particular, if the extracted symbolic program does not satisfy certain desirable verification constraints, unlike previous approaches, our framework allows for repairing the programs in symbolic space and distilling the programs to policies for further optimization.

5 Discussion and Future Work

We presented a preliminary instantiation of the Morl framework showing the benefits of learning a symbolic representation of the policy. Namely, that by optimizing the policy by iterating between two representations, we were able to converge faster to near-optimal performance starting with a poor initialization.

There are a number of assumptions we make in this paper in order to instantiate our framework. While the Morl framework is general enough to encapsulate many different approaches of synthesis, repair, and imitation, we only consider the simplest forms of these. For instance, we hand-design the candidate repaired programs, and use a simple supervised approach for imitation learning. Each of these aspects could be significantly scaled up to be used for larger programs and for more complicated tasks. While CartPole was a simple sandbox for which we could test symbolic programs, for more complicated tasks, automated program repair and verification techniques would be more efficient.

Reward design (Clark & Amodei, 2016) and safety (Hadfield-Menell et al., 2017) is another exciting research direction. Note that we can instead use the reward function as the program representation for Morl; this would instead provide a procedure for more interpretable or verifiable inverse reinforcement learning.


  • Abbeel & Ng (2004) Abbeel, Pieter and Ng, Andrew Y. Apprenticeship learning via inverse reinforcement learning. In Proceedings of the twenty-first international conference on Machine learning, pp.  1. ACM, 2004.
  • Barto et al. (1983) Barto, A. G., Sutton, R. S., and Anderson, C. W. Neuronlike adaptive elements that can solve difficult learning control problems. IEEE Transactions on Systems, Man, and Cybernetics, SMC-13(5):834–846, Sept 1983. ISSN 0018-9472.
  • Bastani et al. (2018) Bastani, Osbert, Pu, Yewen, and Solar-Lezama, Armando. Verifiable reinforcement learning via policy extraction. arXiv preprint arXiv:1805.08328, 2018.
  • Bratko et al. (1995) Bratko, Ivan, Urbančič, Tanja, and Sammut, Claude. Behavioural cloning: phenomena, results and problems. IFAC Proceedings Volumes, 28(21):143–149, 1995.
  • Brockman et al. (2016) Brockman, Greg, Cheung, Vicki, Pettersson, Ludwig, Schneider, Jonas, Schulman, John, Tang, Jie, and Zaremba, Wojciech. Openai gym. arXiv preprint arXiv:1606.01540, 2016.
  • Clark & Amodei (2016) Clark, Jack and Amodei, Dario. Faulty reward functions in the wild., 2016.
  • Hadfield-Menell et al. (2017) Hadfield-Menell, Dylan, Milli, Smitha, Abbeel, Pieter, Russell, Stuart J, and Dragan, Anca. Inverse reward design. In Advances in Neural Information Processing Systems, pp. 6768–6777, 2017.
  • Ho & Ermon (2016) Ho, Jonathan and Ermon, Stefano. Generative adversarial imitation learning. In Advances in Neural Information Processing Systems, pp. 4565–4573, 2016.
  • Jobstmann et al. (2005) Jobstmann, Barbara, Griesmayer, Andreas, and Bloem, Roderick. Program repair as a game. In CAV, pp. 226–238, Berlin, Heidelberg, 2005. Springer-Verlag. doi: 10.1007/11513988˙23. URL
  • Mnih et al. (2015) Mnih, Volodymyr, Kavukcuoglu, Koray, Silver, David, Rusu, Andrei A., Veness, Joel, Bellemare, Marc G., Graves, Alex, Riedmiller, Martin A., Fidjeland, Andreas, Ostrovski, Georg, Petersen, Stig, Beattie, Charles, Sadik, Amir, Antonoglou, Ioannis, King, Helen, Kumaran, Dharshan, Wierstra, Daan, Legg, Shane, and Hassabis, Demis. Human-level control through deep reinforcement learning. Nature, 518(7540):529–533, 2015.
  • Ross et al. (2011) Ross, Stéphane, Gordon, Geoffrey, and Bagnell, Drew. A reduction of imitation learning and structured prediction to no-regret online learning. In AISTATS, pp. 627–635, 2011.
  • Rusu et al. (2015) Rusu, Andrei A, Colmenarejo, Sergio Gomez, Gulcehre, Caglar, Desjardins, Guillaume, Kirkpatrick, James, Pascanu, Razvan, Mnih, Volodymyr, Kavukcuoglu, Koray, and Hadsell, Raia. Policy distillation. arXiv preprint arXiv:1511.06295, 2015.
  • Schulman et al. (2015) Schulman, John, Levine, Sergey, Abbeel, Pieter, Jordan, Michael, and Moritz, Philipp. Trust region policy optimization. In International Conference on Machine Learning, pp. 1889–1897, 2015.
  • Schulman et al. (2017) Schulman, John, Wolski, Filip, Dhariwal, Prafulla, Radford, Alec, and Klimov, Oleg. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017.
  • Silver et al. (2016) Silver, David, Huang, Aja, Maddison, Chris J., Guez, Arthur, Sifre, Laurent, van den Driessche, George, Schrittwieser, Julian, Antonoglou, Ioannis, Panneershelvam, Vedavyas, Lanctot, Marc, Dieleman, Sander, Grewe, Dominik, Nham, John, Kalchbrenner, Nal, Sutskever, Ilya, Lillicrap, Timothy P., Leach, Madeleine, Kavukcuoglu, Koray, Graepel, Thore, and Hassabis, Demis. Mastering the game of go with deep neural networks and tree search. Nature, 529(7587):484–489, 2016.
  • Silver et al. (2017) Silver, David, Hubert, Thomas, Schrittwieser, Julian, Antonoglou, Ioannis, Lai, Matthew, Guez, Arthur, Lanctot, Marc, Sifre, Laurent, Kumaran, Dharshan, Graepel, Thore, Lillicrap, Timothy P., Simonyan, Karen, and Hassabis, Demis. Mastering chess and shogi by self-play with a general reinforcement learning algorithm. CoRR, abs/1712.01815, 2017.
  • Singh et al. (2013) Singh, Rishabh, Gulwani, Sumit, and Solar-Lezama, Armando. Automated feedback generation for introductory programming assignments. In PLDI, pp. 15–26, 2013.
  • Verma et al. (2018) Verma, Abhinav, Murali, Vijayaraghavan, Singh, Rishabh, Kohli, Pushmeet, and Chaudhuri, Swarat. Programmatically interpretable reinforcement learning. arXiv preprint arXiv:1804.02477, 2018.
  • Weimer et al. (2009) Weimer, Westley, Nguyen, ThanhVu, Le Goues, Claire, and Forrest, Stephanie. Automatically finding patches using genetic programming. In ICSE, pp. 364–374. IEEE Computer Society, 2009.
Comments 0
Request Comment
You are adding the first comment!
How to quickly get a good reply:
  • Give credit where it’s due by listing out the positive aspects of a paper before getting into which changes should be made.
  • Be specific in your critique, and provide supporting evidence with appropriate references to substantiate general statements.
  • Your comment should inspire ideas to flow and help the author improves the paper.

The better we are at sharing our knowledge with each other, the faster we move forward.
The feedback must be of minimum 40 characters and the title a minimum of 5 characters
Add comment
Loading ...
This is a comment super asjknd jkasnjk adsnkj
The feedback must be of minumum 40 characters
The feedback must be of minumum 40 characters

You are asking your first question!
How to quickly get a good answer:
  • Keep your question short and to the point
  • Check for grammar or spelling errors.
  • Phrase it like a question
Test description