Systems and Methods for Providing Reinforcement Learning in a Deep Learning System

Abstract

Systems and methods for providing reinforcement learning for a deep learning network are disclosed. A reinforcement learning process that provides deep exploration is provided by a bootstrap that applied to a sample of observed and artificial data to facilitate deep exploration via a Thompson sampling approach.

Description

FIELD OF THE INVENTION

This invention relates to deep learning networks including, but not limited to, artificial neural networks. More particularly, this invention relates to systems and methods for training deep learning networks from a set of training data using reinforcement learning.

BACKGROUND OF THE INVENTION

Deep learning networks including, but not limited to, artificial neural networks are machine learning systems that receive data, extract statistics and classify results. These systems use a training set of data to generate a model in order to make data driven decisions to provide a desired output.

Deep learning networks are often used to solve problems in an unknown environment where the training dataset is used to ascertain the extent of the environment. One manner of training a deep learning network is referred to as reinforcement learning in which the system takes a sequence of actions in order to maximize cumulative rewards. In reinforcement learning, the system begins with an imperfect knowledge of the environment and learns through experience. As such, there is a fundamental trade-off between exploration and exploitation in that the system may improve its future rewards by exploring poorly understood states and actions and sacrificing immediate rewards.

Many approaches to reinforcement learning have been put forth. Most of the proposed approaches are designed based upon Markov Decision Processes (MDPs) with small finite state spaces. Some other approaches require solving computationally intractable planning tasks. These approaches are not practical in complex environments that require the system to generalize in order to operate properly. Thus, these reinforcement learning approaches in large-scale application have relied upon either statistically inefficient exploration strategies or include no exploration at all.

Some other common exploration approaches are dithering strategies. An example of a dithering strategy is a &greedy. In common dithering strategies, the approximated value of an action is a single value and the system picks the action with the highest value. In some strategies, the system may also choose some actions at random. Another common exploration strategy is inspired by Thompson sampling. In a Thompson sampling strategy, there is some notion of uncertainty. However, a distribution of the possible values from the dataset and the system is explored by randomly selecting a policy according to the probability that the selected policy is the optimal policy.

SUMMARY

The above and other problems are solved and an advance in the art is made by systems and methods for providing reinforcement learning in deep learning networks in accordance with some embodiments of the invention. Reinforcement learning with deep exploration is provided in the following manner in accordance with some embodiments of the invention. A deep neural network is maintained. A reinforcement learning process is applied to the deep neural network.

The reinforcement learning process is performed in the following manner in accordance with some embodiments. A set of observed data and a set artificial data is received. For each of a number of episodes, the process samples a set of data that is a union of the set of observed data and the set of artificial data to generate set of training data. A state-action value function is then determined for the set of training data using a bootstrap process and an approximator. The approximator estimates a state-action function for a dataset. For each time step in each of the one or more episodes, the process determines a state of the system for a current time step from the set of training data. An action based on the determined state of the system and a policy mapping actions to the state of the system is selected by the process and results for the action including a reward and a transition state that result from the selected action are determined. Result data from the current time step that includes the state, the action, the transition state are stored. The set of the observed data is then updated with the result data from at least one time step of an episode at the conclusion of an episode.

In accordance with some embodiments, the reinforcement learning process generates the set of artificial data from the set of observed data. In accordance with many of these embodiments the artificial data is generated by sampling the set of observed data with replacement to generate the set of artificial data. In accordance with a number of other embodiments, the artificial data is generated by sampling state-action pairs from a diffusely mixed generative model and assigning each of the sampled state-action pairs stochastically optimistic rewards and random state transitions.

In accordance with some embodiments, the reinforcement learning process maintains a training mask that indicates the result data from each of the time period in each episode to be used in training and updates the set of observed data by adding the result data from each time period of an episode indicated in the training mask.

In accordance with some embodiments, the approximator can be received as an input. In accordance with many embodiments, the approximator is read from memory. In accordance with a number of embodiments, the approximator is a neural network trained to fit a state-action value function to the data set via a least squared iteration.

In accordance with some embodiments, one or more reinforcement learning processes are applied to the deep neural network. In accordance with many of these embodiments, each of the reinforcement learning processes independently maintains a set of observed data. In accordance with some other embodiments, the reinforcement learning processes cooperatively maintain the set of observed data. In accordance with some of these embodiments, a bootstrap mask that indicates each element in the set of observed data that is available to each of the reinforcement learning processes is maintained.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 illustrates a process for a bootstrap method in accordance with an embodiment of the invention.

FIG. 2 illustrates a process for a Bayesian bootstrap method in accordance with an embodiment of the invention.

FIG. 3 illustrates a bootstrapped Thompson sampling process in accordance with an embodiment of the invention.

FIG. 4 illustrates plots for different variants of bootstrapped Thompson sampling in accordance with an embodiment of the invention.

FIG. 5 illustrates a process for reinforcement learning with bootstrapped value function randomization in accordance with an embodiment of the invention.

FIG. 6 illustrates a process for incremental fitting by leveraging parallel computation in accordance with an embodiment of the invention.

FIG. 7 illustrates a process for incremental fitting by leveraging parallel computation in accordance with an embodiment of the invention.

FIG. 8 illustrates various devices in a network that perform processes that systems and methods for providing reinforcement learning in a deep learning network in accordance with various embodiments of the invention.

FIG. 9 illustrates a processing system in a device that performs processes that provide systems and methods for providing reinforcement learning in a deep learning network in accordance with various embodiments of the invention.

FIG. 10 illustrates a process for providing reinforcement learning in a deep learning network in accordance with an embodiment of the invention.

FIG. 11 illustrates an incremental process for providing reinforcement learning in a deep learning network in accordance with an embodiment of the invention.

DETAILED DISCUSSION

Turning now to the drawings, systems and methods for providing reinforcement learning to a deep learning network in accordance with various embodiment of the invention are disclosed. For purposes of this discussion, deep learning networks are machine learning systems that use a dataset of observed data to learn how to solve a problem in a system where all of the states of the system, actions based upon states, and/or the resulting transitions are not fully known. Examples of deep learning networks include, but are not limited to, deep neural networks.

System and methods in accordance with some embodiments of this invention that provide reinforcement learning do so by providing an exploration process for a deep learning network to solve a problem in an environment. In reinforcement learning, actions taken by a system may impose delayed consequences. Thus, the design of exploration strategies is more difficult than systems that are action-response systems where there are no delayed consequences such as multi-armed bandit problems because the system must establish a context. An example of a system that has delayed consequences is a system that interacts with an environment over repeated episodes, I, of length τ. In each time step, t=1, . . . τ, of an episode, the system observes a state of the environment, i_lt, and selects an action, a_lt, according to a policy π which maps the states to actions. A reward, r_lt, and a state transition to state, s_lt+1, are realized in response to the action. The goal of the system during exploration is to maximize the long-term sum of the expected rewards even though the system is unsure of the dynamics of the environment and the reward structure.

To perform well in a sequential decision task while learning about its environment, an agent must balance between exploitation, making good decisions given available data, and exploration, taking actions that may help to improve the quality of future decisions. Upper-confidence bound approaches offer one popular class of exploration heuristics that come with performance guarantees. Such approaches assign to poorly-understood actions high but statistically plausible values, effectively allocating an optimism bonus to incentivize exploration of each such action. However, designing, tuning, and applying such algorithms can be challenging or intractable, and as such, upper-confidence bound algorithms applied in practice often suffer poor empirical performance.

Another popular heuristic, which on the surface appears unrelated, is called Thompson sampling or probability matching. In this algorithm the agent maintains a posterior distribution of beliefs and, in each time period, samples an action randomly according to the probability that it is optimal. Although this algorithm is one of the oldest recorded exploration heuristics, it received relatively little attention until recently when its strong empirical performance was noted, and a host of analytic guarantees followed. It turns out that there is a deep connection between Thompson sampling and optimistic algorithms; in particular, Thompson sampling can be viewed as a randomized approximation that approaches the performance of a well-designed and well-tuned upper confidence bound algorithm.

Almost all of the literature on Thompson sampling takes the ability to sample from a posterior distribution as given. For many commonly used distributions, this is served through conjugate updates or Markov chain Monte Carlo methods. However, such methods do not adequately accommodate contexts in which models are nonlinearly parameterized in potentially complex ways, as is the case in deep learning. We introduce an alternative approach to tractably attain the behavior of Thompson sampling in a manner that accommodates such nonlinearly parameterized models and, in fact, may offer advantages more broadly when it comes to efficiently implementing and applying Thompson sampling.

The approach we propose is based on a bootstrap technique that uses a combination of observed and artificially generated data. The idea of using the bootstrap to approximate a posterior distribution is not new, and has been noted from inception of the bootstrap concept. Further, the application of the bootstrap and other related sub-sampling approaches to approximate Thompson sampling is not new. However, we show that these existing approaches fail to ensure sufficient exploration for effective performance in sequential decision problems. As we will demonstrate, the way in which we generate and use artificial data is critical. The approach is particularly well-suited for contexts in which exploration is coupled with deep learning, since in such settings, maintaining or generating samples from a posterior distribution becomes computationally infeasible. Further, our approach is parallelizable and as such scales well to massive complex problems. We explain how the approach can be applied to multi-armed bandit and reinforcement learning problems and how it relates to Thompson sampling.

The term bootstrap refers to a class of methods for nonparametric estimation from data-driven simulation. In essence, the bootstrap uses the empirical distribution of a sampled dataset as an estimate of the population statistic. FIG. 1 provides pseudocode for what is perhaps the most common form of bootstrap. We use P(X) to denote the set of probability measures over a set X.

The procedure of FIG. 1 allows estimates of the distribution of any unknown parameter in a non-parametric manner. The procedure of FIG. 1 easily specializes to familiar concrete versions. For example, suppose ϕ is the expectation operator. Then, each sample y_k is the mean of a sample of the data set, and {circumflex over (P)} is the relative frequency measure of these sample means.

The output {circumflex over (P)} is reminiscent of a Bayesian posterior with an extremely weak prior. For the Bayesian bootstrap as illustrated in FIG. 2, the distribution produced can be interpreted as a posterior based on the data and a degenerate Dirichlet prior.

With the bootstrap approaches we have described, the support of distributions {circumflex over (P)}_k is restricted to the dataset {x₁, . . . , x_N}. We will show that in sequential decision problems this poses a significant problem. To address the problem, we propose a simple extension to the bootstrap. In particular, we augment the dataset {x₁, . . . , x_N} with artificially generated samples {x_N+1, . . . , x_N+M} and apply the bootstrap to the combined dataset. The artificially generated data can be viewed as inducing a prior distribution. In fact, if X is finite and {x_N+1, . . . ,x_N+M}=X, then as K grows, the distribution {circumflex over (P)} produced by the Bayesian bootstrap converges to the posterior distribution conditioned on {x₁, . . . , x_N}, given a uniform Dirichlet prior.

When the set X is large or infinite, the idea of generating one artificial sample per element does not scale gracefully. If we require one prior observation for every single possible value then the observed data {x₁, . . . ,x_N} will bear little influence on the distribution {circumflex over (P)} relative to the artificial data {x_N+1, . . . , x_N+M}. To address this, for a selected value of M, we sample the M artificial data points x_N+1, . . . , X_N+M from a “prior” distribution P₀. The important thing here is that the relative strength M/N of the induced prior can be controlled in an explicit manner. Similarly, depending on the choice of the prior sampling distribution P₀ the posterior is no longer restricted to finite support. In many ways this extension corresponds to using a Dirichlet process prior with generator P₀.

This augmented bootstrap procedure is especially promising for nonlinear functions such as deep neural networks, where sampling from a posterior is typically intractable. In fact, given any method of training a neural network on a dataset of size N, we can generate an approximate posterior through K bootstrapped versions of the neural network. In its most naive implementation this increases the computational cost by a factor of K. However, this approach is parallelizable and therefore scalable with compute power. In addition, it may be possible to significantly reduce the computational cost of this bootstrap procedure by sharing some of the lower-level features between bootstrap sampled networks or growing them out in a tree structure. This could even be implemented on a single chip through a specially constructed dropout mask for each bootstrap sample. With a deep neural network this could provide significant savings.

Multi-armed Bandit

Consider a problem in which an agent sequentially chooses actions (A_t: t ∈ ) from an action set A and observes corresponding outcomes (Y_t,At: t ∈ ). There is a random outcome Y_t,At ∈ Y associated with each α ∈ A and time t ∈ . For each random outcome the agent receives a reward R(Y_t,a) where R:Y is a known function. The “true outcome distribution” p* is itself drawn from a family of distributions P. We assume that, conditioned on p*, (Y_t: t ∈ ) is an iid sequence with each element Y_t distributed according to p*. Let p_a^* be the marginal distribution corresponding to Y_t,a. The T-period regret of the sequence of actions A₁, . . . , A_T is the random variable

$\mathrm{Regret}\left(T,p^{*}\right)=\sum _{t=1}^T\left[\underset{a}{\max }R\left(Y_{t,a}\right)-R\left(Y_{t,A_t}\right)|p^{*}\right]$

The Bayesian regret to time T is defined by BayesRegret(T)=[Regret(T,p*)], where the expectation is taken with respect to the prior distribution over p*.

We take all random variables to be defined with respect to a probability space (Ω, F, ). We will denote by H_t the history of observations (A₁,Y_1,A1, . . . , A_t−1,Y_t−1,At−1) realized prior to time t. Each action A_t is selected based only on H_t and possibly some external source of randomness. To represent this external source, we introduce a sequence of iid random variables (U_t: t ∈ ). Each action A_t is measurable with respect to the sigma-algebra generated by (H_t, U_t).

The objective is to choose actions in a manner that minimizes Bayesian regret. For this purpose, it is useful to think of actions as being selected by a randomized policy π=(π_t: t ∈ ), where each in is a distribution over actions and is measurable with respect to the sigma-algebra generated by H_t. An action A_t is chosen at time t by randomizing according to π_t(·)=(A_t ∈·|H_t). Our bootstrapped Thompson sampling, presented in FIG. 3, serves as such a policy. The algorithm uses a bootstrap algorithm, like those described with reference to FIGS. 1 and 2, as a subroutine. Note that the sequence of action-observation pairs is not iid, though bootstrap algorithms effectively treat data passed to them as iid. The function [p*|H_t+m=·] passed to the bootstrap algorithm maps a specified history of t+M action-observation pairs to a probability distribution over reward vectors. The resulting probability distribution can be thought of as a model fit to the data provided in the history. Note that the algorithm takes as input a distribution {tilde over (P)} from which artificial samples are drawn. This can be thought of as a subroutine that generates M action-observation pairs. As a special case, this subroutine could generate M deterministic pairs. There can be advantages, though, to using a stochastic sampling routine, especially when the space of action-observation pairs is large and we do not want to impose too strong a prior.

The process illustrated in FIG. 3 is similar to Thompson sampling, though the posterior sampling step has been replaced by a single bootstrap sample. For several multi-armed bandit problems BootstrapThompson with appropriate artificial data is equivalent to Thompson sampling.

One drawback of the process illustrated in FIG. 3 is that the computational cost per timestep grows with the amount of data H_t. For applications at large scale this will be prohibitive. Fortunately there is an effective method to approximate these bootstrap samples in an online manner at constant computational cost. Instead of generating a new bootstrap sample every step we can approximate the bootstrap bootstrap distribution by training D ∈ online bootstrap models in parallel and then sampling between them uniformly. In its most naive implementation this parrallel bootstrap will have a computational cost per timestep D times larger than a greedy algorithm. However, for specific function classes such as neural networks it may be possible to share some computation between models and provide significant savings.

We now examine a simple problem instance designed to demonstrate the need for an artificial history to incentivize efficient exploration in BootstrapThompson. The action space A={1, 2}, outcomes Y=[0, 1] and rewards R(y)=y. We fix 0<∈<<1 and describe the true underlying distribution in terms of the Dirac delta function δ_x(y) which assigns all probability mass to x:

$p_a^{*}\left(y\right)=\{\begin{array}{cc}{\delta }_{ϵ}\left(y\right)& \mathrm{if}a=1\\ \left(1-2ϵ\right){\delta }_0\left(y\right)+2ϵ{\delta }_1\left(y\right)& \mathrm{if}a=2\end{array}$

The optimal policy is to pick a=2 at every timestep, since this has an expected reward of 2 ∈ instead of just ∈. However, with probability at least 1-2 ∈, BootstrapThompson without artificial history (M=0) will never learn the optimal policy.

To see why this is the case, note that BootstrapThompson without artificial history must begin by sampling each arm once. In our system this means that with probability 1-2 ∈ the agent will receive a reward of ∈ from arm one and 0 from arm two. Given this history, the algorithm will prefer to choose arm one for all subsequent timesteps, since its bootstrap estimates will always put all their mass on ∈ and 0 respectively. However, we show that this failing can easily be remedied by the inclusion of some artificial history.

In FIG. 4, we plot the cumulative regret of three variants of BootstrapThompson using different bootstrap algorithms with M=0 or M=2. For our simulations we set ∈=0.01 and ran 20 Monte Carlo simulations for each variant of the algorithm. In each simulation and each time period, the pair of artificial data points for cases with M=2 was sampled from a distribution {tilde over (P)} that selects each of the two actions once and samples an observation uniformly from [0, 1] for each. We found that, in this example, the choice of bootstrap method makes little difference but that injecting artificial data is crucial to incentivizing efficient exploration.

The six subplots in FIG. 4 present results differentiated by choice of bootstrap algorithms and whether or not to include artificial data. The columns indicate which bootstrap algorithm is used with labels “ Bootstrap” for a process like that described in FIG. 1, “Bayes” for a process like that described in FIG. 2, and “BESA” for a recently proposed bootstrap approach. The BESA algorithm is a variant of the bootstrap that applies to two armed bandit problems. In each time period, the algorithm estimates the reward of each arm by drawing a sample average (with replacement) with sample size equal to the number of times the other arm has been played.

The rows indicate whether or not artificial prior data was used. We see that BootstrapThompson generally fails to learn with M=0, however the inclusion of artificial data helps to drive efficient exploration. The choice of bootstrap algorithm B seems to make relatively little difference, however we do find that Bootstrap and Bayes seem to outperform BESA on this example.

The BootstrapThompson algorithm is similar to Thompson sampling, the only difference being that a draw from the posterior distribution is replaced by a bootstrap sample. In fact, we can show that, for particular choices of bootstrap algorithm and artificial data distribution, the two algorithms are equivalent.

Consider as an example a multi-armed bandit problem with independent arms, for which each ath arm generates rewards from a Bernoulli distribution with mean θ_a. Suppose that our prior distribution over θ_a is Beta(α_a,β_a). Then, the posterior conditioned on observing n_a0 outcomes with reward zero and n_a1 outcomes with reward one is Beta(α_a+n_a0,β_a+n_a1). If α_a and β_a are positive integers, a sample {circumflex over (θ)}_a from this distribution can be generated by the following procedure: sample x₁, . . . , x_αa+na1,y₁, . . . , y_βa+na2˜Exp (1) and let {circumflex over (θ)}a=Σ_ix_i/(Σ_ix_i+Σ_jy_j). This sampling procedure is identical to the procedure described with reference to FIG. 2, with artificial data generated by a distribution {tilde over (P)} that assigns all probability to a single outcome that, for each arm a , produces α_a+β_a data samples, with α_a of them associated with reward one and β_a of them associated with reward 0.

The example we have presented can easily be generalized to the case where each arm generates rewards from among a finite set of possibilities with probabilities distributed according to a Dirichlet prior. We expect that, with appropriately designed schemes for generating artificial data, such equivalences can also be established for a far broader range of problems.

The aforementioned equivalencies imply that theoretical regret bounds previously developed for Thompson sampling apply to the BootstrapThompson algorithm with the Bayesian bootstrap and appropriately generated artificial data.

Reinforcement Learning

Reinforcement learning is a deep learning approach that differs from standard supervised learning in that correct input/output pairs are never presented, nor sub-optimal actions explicitly corrected. Further, there is a focus on on-line performance, which involves finding a balance between exploration (of uncharted territory) and exploitation (of current knowledge).

In reinforcement learning, actions taken by the agent can impose delayed consequences. This makes the design of exploration strategies more challenging than for multi-armed bandit problems. To fix a context, consider an agent that interacts with an environment over repeated episodes of length π. In each time period t=1, . . . , π of each episode l=1, 2, . . . , the agent observes a state s_lt and selects an action a_lt according to a policy π which maps states to actions. A reward of r_lt and state transition to s_lt+1 are then realized. The agent's goal is to maximize the long term sum of expected rewards, even though she is initially unsure of the system dynamics and reward structure.

A common approach to reinforcement learning involves learning a state-action value function, Q, which for time, t, state, s, and action, a, provides an estimate, Q_t(s, a), of expected rewards over the remainder of the episode: r_l,t+r_l,t+1+ . . . +r_l,π. Given a state-action value function, Q, the system selects an action that maximizes Q_t(s,a) when at state s and time t. Most reinforcement learning systems provide an exploration strategy that balances exploration with exploitation. However, the vast majority of these processes operate in a “tabula rasa” setting which does not allow for generalization between state-action pairs which is needed in systems having a large number of states and actions.

Some of the most successful applications of reinforcement learning generalize using nonlinearly parameterized models, like deep neural networks, that approximate the state-action value function. These algorithms have attained superhuman performance and generated excitement for a new wave of artificial intelligence, but still fail at simple tasks that require efficient exploration since they use simple exploration schemes that do not adequately account for the possibility of delayed consequences. Recent research has shown how to combine efficient generalization and exploration via randomized linearly parameterized value functions. Some approaches can be viewed as a variant of Thompson sampling applied in a reinforcement learning context. But such approaches do not serve the needs of nonlinear parameterizations. A version of Thompson sampling that serves such needs via leveraging the bootstrap and artificial data is illustrated in FIG. 5.

Before the beginning of each episode, the algorithm applies cp to generate a randomized state-action value function. The agent then follows the greedy policy with respect to that sample over the entire episode. As is more broadly the case with Thompson sampling, the algorithm balances exploration with exploitation through the randomness of these samples. The algorithm enjoys the benefits of what we call deep exploration in that it sometimes selects actions which are neither exploitative nor informative in themselves, but that are oriented toward positioning the agent to gain useful information downstream in the episode. In fact, the general approach represented by this algorithm may be the only known computationally efficient means of achieving deep exploration with nonlinearly parameterized representations such as deep neural networks.

As discussed earlier, the inclusion of an artificial history can be crucial to incentivize proper exploration in multi-armed bandit problems. The same is true for reinforcement learning. One simple approach to generating artificial data that accomplishes this in the context of a process such as that described in FIG. 5, is to sample state-action pairs from a diffusely mixed generative model and assign them stochastically optimistic rewards and random state transitions. When prior data is available from episodes of experience with actions selected by an expert agent, one can also augment this artificial data with that history of experience. This offers a means of incorporating apprenticeship learning as a springboard for the learning process.

Fitting a model like a deep neural network can itself be a computationally expensive task. As such it is desirable to use incremental methods that incorporate new data samples into the fitting process as they appear, without having to refit the model from scratch. It is important to observe that a slight variation of the process described in FIG. 5 accommodates this sort of incremental fitting by leveraging parallel computation. This variation is presented in FIGS. 6 and 7. The process makes use of an incremental model learning method ϕ, which takes as input a current model, previous data set, and new data point, with a weight assigned to each data point. The process of FIG. 7 maintains κ models (for example, κ deep neural networks), incrementally updating each in parallel after observing each episode. The model used to guide action in an episode is sampled uniformly from the set of κ. It is worth noting that this is akin to training each model using experience replay, but with past experiences weighted randomly to induce exploration.

A reinforcement learning system in accordance with embodiments of this invention overcomes these problems by providing an exploration strategy that combines efficient generalization and exploration via leveraging a bootstrap process and artificial data. In accordance with some embodiments of the invention, the system receives a set of training data that includes observed data and artificial data. The latter serves to induce a prior distribution which, as we will demonstrate, is critical to effective exploration. In accordance with some embodiments, the artificial data is generated by sample state-action pairs from a diffusely mixed generative model and assign each state-action pair stochastically optimistic rewards and random state transitions. In accordance with some other embodiments, the artificial data is generated by sampling a set of observed data with replacement to obtain a set of data having a number of elements that is approximately equal to or greater than the number of elements as the set of observed data.

The observed and artificial data are sampled to obtain a training sample set of data. In accordance with some of these embodiments, the training sample dataset includes M samples of data. In accordance with a number of these embodiments, M is equal to or greater than a number of episodes to observe during an exploration process. In accordance with some embodiments, the sampling is performed in accordance with a known and/or a provided distribution. A bootstrap process is applied to the union of the observed data and the artificial data to obtain a new distribution for the sample of data. An approximator function is applied to the new distribution to generate a randomized state-value function.

For each episode, the process observes a state of the system, s_lt, for a particular time period from the training sample dataset and selects an action to perform based upon policy π. The results including a reward, r_lt realized and a resulting transition state, s_lt+1 resulting from the action are observed. The state, action, reward and transition state are stored as result data for the episode. This is repeated for each time step in the episode. After the episode is completed, the observed dataset is updated with the result data stored during the episode. In accordance with some embodiments, a training mask may be maintained that includes flags indicating the result data from particular time steps is to be used for training. The training mask is then read to determine the result data to add to the observed data.

In accordance with many embodiments, multiple exploration processes may be performed concurrently. In accordance with a number of these embodiments, the result data observed from each exploration process is shared with other exploration processes. In accordance with a number of embodiments, the observed dataset for each process is updated independently. In accordance with some of these embodiments, a boot strap mask is maintained that indicates which elements of the observed dataset are available to each process.

Systems and method for providing reinforcement learning in a deep learning network in accordance with some embodiments of the invention are set forth below with reference to the Figures.

Systems that Provide Deep Learning Neworks

A system that provides a deep learning system that uses systems and methods that provide reinforcement learning in accordance with some embodiments of the invention is shown in FIG. 8. Network 800 includes a communications network 860. The communications network 860 is a network such as the Internet that allows devices connected to the network 860 to communicate with other connected devices. Server systems 810, 840, and 870 are connected to the network 860. Each of the server system 810, 840, and 870 is a group of one or more servers communicatively connected to one another via internal networks that execute processes that provide cloud services to users over the network 860. For purposes of this discussion, cloud services are one or more applications that are executed by one or more server systems to provide data and/or executable applications to devices over a network. The server systems 810, 840, and 870 are shown each having three servers in the internal network. However, the server systems 810, 840 and 870 may include any number of servers and any additional number of server systems may be connected to the network 860 to provide cloud services. In accordance with various embodiments of this invention, a deep learning network that uses systems and methods that provide reinforcement learning in accordance with an embodiment of the invention may be provided by process being executed on a single server system and/or a group of server systems communicating over network 860.

Users may use personal devices 880 and 820 that connect to the network 860 to perform processes for providing and/or interaction with a deep learning network that uses systems and methods that provide reinforcement learning in accordance with various embodiments of the invention. In the shown embodiment, the personal devices 880 are shown as desktop computers that are connected via a conventional “wired” connection to the network 860. However, the personal device 880 may be a desktop computer, a laptop computer, a smart television, an entertainment gaming console, or any other device that connects to the network 860 via a “wired” connection. The mobile device 820 connects to network 860 using a wireless connection. A wireless connection is a connection that uses Radio Frequency (RF) signals, Infrared signals, or any other form of wireless signaling to connect to the network 860. In FIG. 8, the mobile device 820 is a mobile telephone. However, mobile device 820 may be a mobile phone, Personal Digital Assistant (PDA), a tablet, a smartphone, or any other type of device that connects to network 860 via wireless connection without departing from this invention.

Example of a Processing System

An example of a processing system in a device that executes instructions to perform processes that provide interact with other devices connected to the network as shown in FIG. 8 to provide a deep learning network that uses systems and methods that provide reinforcement learning in accordance with various embodiments of the invention in accordance with various embodiments of this invention is shown in FIG. 9. One skilled in the art will recognize that a particular processing system may include other components that are omitted for brevity without departing from this invention. The processing device 900 includes a processor 905, a non-volatile memory 910, and a volatile memory 915. The processor 905 is a processor, microprocessor, controller, or a combination of processors, microprocessor, and/or controllers that performs instructions stored in the volatile 915 or the non-volatile memory 910 to manipulate data stored in the memory. The non-volatile memory 910 can store the processor instructions utilized to configure the processing system 900 to perform processes including processes in accordance with embodiments of the invention and/or data for the processes being utilized. In other embodiments, the processing system software and/or firmware can be stored in any of a variety of non-transient computer readable media appropriate to a specific application. A network interface is a device that allows processing system 900 to transmit and receive data over a network based upon the instructions performed by processor 905. Although a processing system 900 is illustrated in FIG. 9, any of a variety of processing system in the various devices can configured to provide the methods and systems in accordance with embodiments of the invention can be utilized.

Reinforcement Learning Exploration Process

In accordance with some embodiments of the invention, systems and methods provide reinforcement learning by providing a deep exploration process. The deep exploration process fits a state-action value function to a sample of data from set of data that includes artificial data and observed data. In accordance with some embodiments of the invention, the system receives a set of training data that includes observed data and artificial data. In accordance with some these embodiments, the artificial data is generated by sampling state-action pairs from a diffusely mixed generative model and assign each state-action pair stochastically optimistic rewards and random state transitions. In accordance with some other embodiments, the artificial set of data is generated by sampling the observed set of data with replacement. In accordance with a number of embodiments, the artificial data includes M elements where M is approximately greater than or equal to the number of elements in the observed dataset. The use of the combination of observed and historical data provides randomness in the samples to induce deep learning. An exploration process for providing reinforcement learning to a deep learning network in accordance with an embodiment of this invention is shown in FIG. 10.

Process 1000 performs exploration in M distinct episodes. In accordance with some embodiments, the number of episodes is received as an input and in accordance with some other embodiments the number of episodes may be set or selected by the process based on the size of the deep learning network. A set of data including historical and artificial data is obtained (1005). In the context of our episodic setting, each element of the data set corresponds to a sequence of observations made over an episode. In accordance with some embodiments, the observed data is read from a memory. In accordance with some other embodiments, the observed data is received from another system.

In accordance with some embodiments, the artificial data is generated by sampling the observed data with replacement. In accordance with some other embodiments, the artificial data is generated from the observed data based on a known distribution of the original data. In accordance with some other embodiments, the artificial data is generated independent of the observed data. In accordance with some of these embodiments, the artificial data is generated by sampling state-action pairs from a diffusely mixed generative model; and assigning each state-action pair stochastically optimistic rewards and random state transitions. In accordance with some embodiments, the artificial dataset includes M elements of data where M is approximately equal to or greater than the number of elements in the observed dataset.

An approximator function is also received as input (1010). In accordance with some embodiments, the approximator function may be set for process 1000 and stored in memory for use. The approximator estimates a state-action value function for a data set. In accordance with some embodiments, the approximator function may be a neural network trained to fit a state-action value function to the data set via a least squared iteration. A number of conventional reinforcement learning algorithms would fit the state-action value function to the observed history H. Two key features that distinguish processes in accordance with various embodiments of the invention are that the state-action value function is fit to a random subsample of data and that this subsample is drawn from a combination of historical and artificial data.

The observed and artificial data are sampled to obtain a training set of data (1015). In accordance with some of these embodiments, the training data includes M samples of data. In accordance with a number of these embodiments, M is equal to or greater than a number of episodes to observe. In accordance with some embodiments, the sampling is performed in accordance with a known and/or a provided distribution. A bootstrap process is applied to the union of the observed data and the training data to obtain a new distribution and the approximator function is applied to the distribution to generate a randomized state-value function (1020). For each time step, a state, s, is observed based on the training data and an action, a, is selected based on the state of the system from the sample of data and the policy π (1030). The reward, r_lt realized and resulting transition state, s_lt+1 are observed (1035). The state, s, the action, a, and the resulting transition state, s_lt+1 are stored as resulting data in memory. The selecting (1030) and observing of the results are repeated until the time period ends (1040). The observed set of data is then updated with the results (1050). In accordance with some embodiments, a training mask is maintained that indicates the result data from particular time steps of each episode to add to the observed set of data and the mask is read to determine which elements of the resulting data to add to the observed data. This is then repeated for each of the M episodes (1045) and process 1000 ends.

Although one process for providing reinforcement learning for a deep learning network in accordance with an embodiment of the invention is described with respect to FIG. 10. Other methods that add, removed, and/or combine steps in process 1000 may be performed without departing from the various embodiments of this invention.

Fitting a model like a deep neural network is a computationally expensive task. As such, it is desirable to use incremental methods to incorporate new data sample into the fitting process as the data is generated. To do so, parallel computing may be used. A process that performs multiple concurrent explorations in accordance with an embodiment of the invention is shown in FIG. 11.

Process 1100 performs exploration in M distinct episodes for K separate exploration process. In accordance with some embodiments, the number of episodes is received as an input and in accordance with some other embodiments the number of episodes may set or selected by the process based on the size of the deep learning network. A set of data including historical and artificial data is obtained (1105). In accordance with some embodiments, the observed data is read from a memory. In accordance with some other embodiments, the observed data is received from another system.

In accordance with some embodiments, the artificial data for one or more exploration processes is generated by sampling the observed data with replacement. In accordance with some other embodiments, the artificial data for one or more exploration processes is generated from the observed data based on a known distribution of the original data. In accordance with some other embodiments, the artificial data for one or more of the exploration processes is generated independent of the observed data. In accordance with some of these embodiments, the artificial data is generated by sampling state-action pairs from a diffusely mixed generative model; and assigning each state-action pair stochastically optimistic rewards and random state transitions. In accordance with some embodiments, the artificial dataset includes M elements of data where M is approximately equal to or greater than the number of elements in the observed dataset.

An approximator function is also received as input (1105). In accordance with some embodiments, the approximator function may be set for process 1100 and stored in memory for use. The approximator estimates a state-action value function for a data set. In accordance with some embodiments, the approximator function may be a neural network trained to fit a state-action value function to the data set via a least squared iteration. In accordance with some embodiments, 1 to K number of approximators may be used where each of the 1 to K approximators is applied to the training set data of one or more of the K exploration processes.

The observed and artificial data are sampled to obtain a training set of data for each of the K independent processes (1115). In accordance with some of these embodiments, the training data for each exploration process includes M samples of data. In accordance with a number of these embodiments, M is equal to or greater than a number of episodes to observe. In accordance with some embodiments, the sampling for one or exploration processes is performed in accordance with a known and/or a provided distribution. In accordance with some embodiments, one or more of the exploration process may have the same set of artificial data.

For each of the K exploration processes, a bootstrap process is applied to the union of the observed data and the artificial data to obtain a new distribution and the approximator function is applied to the distribution to generate a randomized state-value function (1120). For each exploration process, exploration is performed in the following manner. For each time step, a state, s, is observed and an action, a, is selected based on the state of the system from the sample of data and the policy π (1130). The reward, r_lt realized and resulting transition state, s_lt+1 are observed (1135). The state, s, the action, a, and the resulting transition state, s_lt+1 are stored as resulting data in memory. The selecting (1130) and observing of the results are repeated until the time period for the episode ends (1140). The observed set of data is individually updated for each exploration process with the results (1150). To do so, a bootstrap mask may be maintained to indicate the observed data that is available to each exploration process. In accordance with some other embodiments, the observed data is updated with the data from all of the different K exploration processes. In accordance with some embodiments, a training mask is maintained that indicates the result data from particular time steps of each episode for each exploration process to add to the observed set of data and the mask is read to determine which elements of the resulting data to add to the observed data. This is then repeated for each of the M episodes (1145) and process 1100 ends.

Although one process for providing reinforcement learning for a deep learning network using multiple exploration processes in accordance with an embodiment of the invention is described with respect to FIG. 11. Other methods that add, removed, and/or combine steps in process 1100 may be performed without departing from the various embodiments of this invention.

Although the present invention has been described in certain specific aspects, many additional modifications and variations would be apparent to those skilled in the art. It is therefore to be understood that the present invention can be practiced otherwise than specifically described without departing from the scope and spirit of the present invention. Thus, embodiments of the present invention should be considered in all respects as illustrative and not restrictive. Accordingly, the scope of the invention should be determined not by the embodiments illustrated, but by the appended claims and their equivalents.

Claims

1. A method for a reinforcement learning process, the method comprising: receiving a set of observed data;generating a set artificial data;drawing a bootstrap sample from a set of data that is a union of the set of observed data and the set of artificial data;determining an approximate state-action value function based on the drawn bootstrap sample;selecting an action based on a state of the system and the approximate state-action value function;determining a result and a transition state for the selected action; andupdating the set of observed data with result data that includes the state, the action, the transition state, and the determined result.

2. The method of claim 1, wherein generating the artificial data comprises sampling from the set of observed data with replacement to generate the set of artificial data.

3. The method of claim 1, wherein generating the artificial data comprises: sampling a plurality of state-action pairs from a diffusely mixed generative model; andassigning each of the plurality of sampled state-action pairs stochastically optimistic rewards and random state transitions.

4. The method of claim 1 wherein the approximate state-action value function is a neural network trained to fit a state-action value function to the bootstrap sample via a least squared iteration.

5. The method of claim 1 further comprising: drawing a second bootstrap sample from the set of data; anddetermining a second approximate state-action value function based on the second bootstrap sample, wherein selecting the action is further based on the second approximate state-action value function.

6. The method of claim 1 further comprising: generating a second set of artificial data;drawing a second bootstrap sample from a second set of data that is a union of the set of observed data and the second set of artificial data; anddetermining a second approximate state-action value function based on the second bootstrap sample, wherein selecting the action is further based on the second approximate state-action value function.

7. The method of claim 6, wherein determining the approximate state-action value function and the second approximate state-action value function are performed in parallel.

8. The method of claim 1, wherein drawing a bootstrap sample, determining an approximate state-action value function, selecting an action, and determining a result and a transition state, are performed for a plurality of iterations.

9. The method of claim 1, wherein selecting an action and determining a result are performed for a sequence of time periods, wherein updating the set result data includes the state, the action, the transition state, and the determined result for the sequence of time periods.

10. The method of claim 1 further comprising: maintaining a training mask that indicates the result data from each time period of the sequence of time periods; andwherein the updating of the set of observed data comprises adding the result data from each time period of an episode indicated in the training mask.

11. A deep learning system for a reinforcement learning process, the system comprising: a set of one or more processors;memory accessible by the set of processors;instructions that when read by the set of processors direct the set of processors to: receive a set of observed data;generate a set artificial data;draw a bootstrap sample from a set of data that is a union of the set of observed data and the set of artificial data;determine an approximate state-action value function based on the drawn bootstrap sample;select an action based on a state of the system and the approximate state-action value function;determine a result and a transition state for the selected action; andupdate the set of observed data with result data that includes the state, the action, the transition state, and the determined result.

12. The deep learning system of claim 11, wherein generating the artificial data comprises sampling from the set of observed data with replacement to generate the set of artificial data.

13. The deep learning system of claim 11, wherein generating the artificial data comprises: sampling a plurality of state-action pairs from a diffusely mixed generative model; andassigning each of the plurality of sampled state-action pairs stochastically optimistic rewards and random state transitions.

14. The deep learning system of claim 11, wherein the approximate state-action value function is a neural network trained to fit a state-action value function to the bootstrap sample via a least squared iteration.

15. The deep learning system of claim 11 further comprising instructions that when read by the set of processors direct the set of processors to: draw a second bootstrap sample from the set of data; anddetermine a second approximate state-action value function based on the second bootstrap sample, wherein selecting the action is further based on the second approximate state-action value function.

16. The deep learning system of claim 11 further comprising instructions that when read by the set of processors direct the set of processors to: generate a second set of artificial data;draw a second bootstrap sample from a second set of data that is a union of the set of observed data and the second set of artificial data; anddetermine a second approximate state-action value function based on the second bootstrap sample, wherein selecting the action is further based on the second approximate state-action value function.

17. The deep learning system of claim 16, wherein determining the approximate state-action value function and the second approximate state-action value function are performed in parallel.

18. The deep learning system of claim 11, wherein the instructions to draw a bootstrap sample, determine an approximate state-action value function, select an action, and determine a result and a transition state, are performed for a plurality of iterations.

19. The deep learning system of claim 11, wherein the instructions to select an action and determine a result are performed for a sequence of time periods, wherein updating the set result data includes the state, the action, the transition state, and the determined result for the sequence of time periods.

20. The deep learning system of claim 11, wherein the instructions further direct the at least one processor to: maintain a training mask that indicates the result data from each time period of the sequence of time periods; andwherein the updating of the set of observed data includes adding the result data from each time period of an episode indicated in the training mask.