Reinforcement learning (RL) is an area of machine learning inspired by behaviourist psychology, concerned with how software agents ought to take actions in an environment so as to maximize some notion of cumulative reward. The problem, due to its generality, is studied in many other disciplines, such as game theory, control theory, operations research, information theory, simulation-based optimization, multi-agent systems, swarm intelligence, statistics and genetic algorithms. In the operations research and control literature, reinforcement learning is called approximate dynamic programming, The approach has been studied in the theory of optimal control, though most studies are concerned with the existence of optimal solutions and their characterization, and not with learning or approximation. In economics and game theory, reinforcement learning may be used to explain how equilibrium may arise under bounded rationality.
In machine learning, the environment is typically formulated as a Markov decision process (MDP), as many reinforcement learning algorithms for this context utilize dynamic programming techniques. The main difference between the classical techniques[which?] and reinforcement learning algorithms is that the latter do not need knowledge[vague] about the MDP and they target large MDPs where exact methods become infeasible.
Reinforcement learning differs from standard supervised learning in that correct input/output pairs[clarification needed] are never presented, nor sub-optimal actions explicitly corrected. Instead the focus is on performance,[clarification needed], which involves finding a balance between exploration (of uncharted territory) and exploitation (of current knowledge). The exploration vs. exploitation trade-off has been most thoroughly studied through the multi-armed bandit problem and in finite MDPs.
Basic reinforcement is modeled as a Markov decision process:
Rules are often stochastic. The observation typically involves the scalar, immediate reward associated with the last transition. In many works, the agent is assumed to observe the current environmental state (full observability). If not, the agent has partial observability. Sometimes the set of actions available to the agent is restricted (a zero balance cannot be reduced).
A reinforcement learning agent interacts with its environment in discrete time steps. At each time t, the agent receives an observation , which typically includes the reward . It then chooses an action from the set of available actions, which is subsequently sent to the environment. The environment moves to a new state and the reward associated with the transition is determined. The goal of a reinforcement learning agent is to collect as much reward as possible. The agent can (possibly randomly) choose any action as a function of the history.
When the agent's performance is compared to that of an agent that acts optimally, the difference in performance gives rise to the notion of regret. In order to act near optimally, the agent must reason about the long term consequences of its actions (i.e., maximize future income), although the immediate reward associated with this might be negative.
Thus, reinforcement learning is particularly well-suited to problems that include a long-term versus short-term reward trade-off. It has been applied successfully to various problems, including robot control, elevator scheduling, telecommunications, backgammon, checkers and go (AlphaGo).
Two elements make reinforcement learning powerful: the use of samples to optimize performance and the use of function approximation to deal with large environments. Thanks to these two key components, reinforcement learning can be used in large environments in the following situations:
The first two of these problems could be considered planning problems (since some form of model is available), while the last one could be considered to be a genuine learning problem. However, reinforcement learning converts both planning problems to machine learning problems.
Reinforcement learning requires clever exploration mechanisms. Randomly selecting actions, without reference to an estimated probability distribution, shows poor performance. The case of (small) finite Markov decision processes is relatively well understood. However, due to the lack of algorithms that provably scale well with the number of states (or scale to problems with infinite state spaces), simple exploration methods are the most practical.
One such method is -greedy, when the agent chooses the action that it believes has the best long-term effect with probability , and it chooses an action uniformly at random, otherwise. Here, is a tuning parameter, which is sometimes changed, either according to a fixed schedule (making the agent explore progressively less), or adaptively based on heuristics.
Even if the issue of exploration is disregarded and even if the state was observable (assumed hereafter), the problem remains to use past experience to find out which actions are good.
The agent's action selection is modeled as a map called policy:
The policy map gives the probability of taking action a when in state s.:61 There are also non-probabilistic policies.
Value function is defined as the expected return starting with state , i.e. , and successively following policy . Hence, roughly speaking, the value function estimates "how good" it is to be in a given state.:60
where the random variable denotes the return, and is defined as the sum of future discounted rewards
where is the reward at step , is the discount-rate.
The algorithm must find a policy with maximum expected return. From the theory of MDPs it is known that, without loss of generality, the search can be restricted to the set of so-called stationary policies. A policy is stationary if the action-distribution returned by it depends only on the last state visited (from the observation agent's history). The search can be further restricted to deterministic stationary policies. A deterministic stationary policy deterministically selects actions based on the current state. Since any such policy can be identified with a mapping from the set of states to the set of actions, these policies can be identified with such mappings with no loss of generality.
The brute force approach entails two steps:
One problem with this is that the number of policies can be large, or even infinite. Another is that variance of the returns may be large, which requires many samples to accurately estimate the return of each policy.
These problems can be ameliorated if we assume some structure and allow samples generated from one policy to influence the estimates made for others. The two main approaches for achieving this are value function estimation and direct policy search.
Value function approaches attempt to find a policy that maximizes the return by maintaining a set of estimates of expected returns for some policy (usually either the "current" [on-policy] or the optimal [off-policy] one).
These methods rely on the theory of MDPs, where optimality is defined in a sense that is stronger than the above one: A policy is called optimal if it achieves the best expected return from any initial state (i.e., initial distributions play no role in this definition). Again, an optimal policy can always be found amongst stationary policies.
To define optimality in a formal manner, define the value of a policy by
where stands for the random return associated with following from the initial state .Defining as the maximum possible value of , where is allowed to change,
A policy that achieves these optimal values in each state is called optimal. Clearly, a policy that is optimal in this strong sense is also optimal in the sense that it maximizes the expected return , since , where is a state randomly sampled from the distribution .
Although state-values suffice to define optimality, it is useful to define action-values. Given a state , an action and a policy , the action-value of the pair under is defined by
where now stands for the random return associated with first taking action in state and following , thereafter.
The theory of MDPs states that if is an optimal policy, we act optimally (take the optimal action) by choosing the action from with the highest value at each state, . The action-value function of such an optimal policy () is called the optimal action-value function and is commonly denoted by . In summary, the knowledge of the optimal action-value function alone suffices to know how to act optimally.
Assuming full knowledge of the MDP, the two basic approaches to compute the optimal action-value function are value iteration and policy iteration. Both algorithms compute a sequence of functions () that converge to . Computing these functions involves computing expectations over the whole state-space, which is impractical for all but the smallest (finite) MDPs. In reinforcement learning methods, expectations are approximated by averaging over samples and using function approximation techniques to cope with the need to represent value functions over large state-action spaces.
Monte Carlo methods can be used in an algorithm that mimics policy iteration. Policy iteration consists of two steps: policy evaluation and policy improvement.
Monte Carlo is used in the policy evaluation step. In this step, given a stationary, deterministic policy , the goal is to compute the function values (or a good approximation to them) for all state-action pairs . Assuming (for simplicity) that the MDP is finite, that sufficient memory is available to accommodate the action-values and that the problem is episodic and after each episode a new one starts from some random initial state. Then, the estimate of the value of a given state-action pair can be computed by averaging the sampled returns that originated from over time. Given sufficient time, this procedure can thus construct a precise estimate of the action-value function . This finishes the description of the policy evaluation step.
In the policy improvement step, the next policy is obtained by computing a greedy policy with respect to : Given a state , this new policy returns an action that maximizes . In practice lazy evaluation can defer the computation of the maximizing actions to when they are needed.
Problems with this procedure include:
The first problem is corrected by allowing the procedure to change the policy (at some or all states) before the values settle. This too may be problematic as it might prevent convergence. Most current algorithms do this, giving rise to the class of generalized policy iteration algorithms. Many actor critic methods belong to this category.
The second issue can be corrected by allowing trajectories to contribute to any state-action pair in them. This may also help to some extent with the third problem, although a better solution when returns have high variance is Sutton'stemporal difference (TD) methods that are based on the recursive Bellman equation. Note that the computation in TD methods can be incremental (when after each transition the memory is changed and the transition is thrown away), or batch (when the transitions are batched and the estimates are computed once based on the batch). Batch methods, such as the least-squares temporal difference method, may use the information in the samples better, while incremental methods are the only choice when batch methods are infeasible due to their high computational or memory complexity. Some methods try to combine the two approaches. Methods based on temporal differences also overcome the fourth issue.
In order to address the fifth issue, function approximation methods are used. Linear function approximation starts with a mapping that assigns a finite-dimensional vector to each state-action pair. Then, the action values of a state-action pair are obtained by linearly combining the components of with some weights :
The algorithms then adjust the weights, instead of adjusting the values associated with the individual state-action pairs. Methods based on ideas from nonparametric statistics (which can be seen to construct their own features) have been explored.
The problem with using action-values is that they may need highly precise estimates of the competing action values that can be hard to obtain when the returns are noisy. Though this problem is mitigated to some extent by temporal difference methods. Using the so-called compatible function approximation method compromises generality and efficiency. Another problem specific to TD comes from their reliance on the recursive Bellman equation. Most TD methods have a so-called parameter that can continuously interpolate between Monte Carlo methods that do not rely on the Bellman equations and the basic TD methods that rely entirely on the Bellman equations. This can be effective in palliating this issue.
An alternative method is to search directly in (some subset of) the policy space, in which case the problem becomes a case of stochastic optimization. The two approaches available are gradient-based and gradient-free methods.
Gradient-based methods (policy gradient methods) start with a mapping from a finite-dimensional (parameter) space to the space of policies: given the parameter vector , let denote the policy associated to . Defining the performance function by
under mild conditions this function will be differentiable as a function of the parameter vector . If the gradient of was known, one could use gradient ascent. Since an analytic expression for the gradient is not available, only a noisy estimate is available. Such an estimate can be constructed in many ways, giving rise to algorithms such as Williams' REINFORCE method (which is known as the likelihood ratio method in the simulation-based optimization literature). Policy search methods have been used in the robotics context. Many policy search methods may get stuck in local optima (as they are based on local search).
A large class of methods avoids relying on gradient information.These include simulated annealing, cross-entropy search or methods of evolutionary computation. Many gradient-free methods can achieve (in theory and in the limit) a global optimum. In multiple domains[which?] they have demonstrated performance.
Policy search methods may converge slowly given noisy data. For example, this happens in episodic problems when the trajectories are long and the variance of the returns is large. Value-function based methods that rely on temporal differences might help in this case. In recent years, actor-critic methods have been proposed and performed well on various problems.
Both the asymptotic and finite-sample behavior of most algorithms is well understood. Algorithms with provably good online performance (addressing the exploration issue) are known.
Efficient exploration of large MDPs is largely unexplored (except for the case of bandit problems).[clarification needed] Although finite-time performance bounds appeared for many algorithms, these bounds are expected to be rather loose and thus more work is needed to better understand the relative advantages and limitations.
For incremental algorithms, asymptotic convergence issues have been settled. Temporal-difference-based algorithms converge under a wider set of conditions than was previously possible (for example, when used with arbitrary, smooth function approximation).
Research topics include
Multiagent or distributed reinforcement learning is a topic of interest. Applications are expanding.
Reinforcement learning algorithms such as TD learning are under investigation as a model for dopamine-based learning in the brain. In this model, the dopaminergic projections from the substantia nigra to the basal ganglia function as the prediction error. Reinforcement learning has been used as a part of the model for human skill learning, especially in relation to the interaction between implicit and explicit learning in skill acquisition (the first publication on this application was in 1995-1996).
Since the work on learning ATARI TV games by Google DeepMind , end-to-end reinforcement learning or deep reinforcement learning is garnering attention. This approach extends reinforcement learning to the entire process from sensors to motors by forming it using an artificial neural network especially a deep network without designing state space or action space explicitly. It can minimize the interference from human design and flexible and purposive learning on a huge degree of freedom enables the emergence of game strategy and other necessary functions inside. That is expected to open up the way to artificial general intelligence (AGI).
In inverse reinforcement learning (IRL), no reward function is given. Instead, the reward function is inferred given an observed behavior from an expert. The idea is to mimic the observed behavior, which is often optimal or close to optimal.
In apprenticeship learning, an expert demonstrates the ideal behavior. The system tries to recover the policy directly using observations from the expert.
Most reinforcement learning papers are published at the major machine learning and AI conferences (ICML, NIPS, AAAI, IJCAI, UAI, AI and Statistics) and journals (JAIR, JMLR, Machine learning journal, IEEE T-CIAIG). Some theory papers are published at COLT and ALT. However, many papers appear in robotics conferences (IROS, ICRA) and the "agent" conference AAMAS. Operations researchers publish their papers at the INFORMS conference and, for example, in the Operation Research, and the Mathematics of Operations Research journals. Control researchers publish their papers at the CDC and ACC conferences, or, e.g., in the journals IEEE Transactions on Automatic Control, or Automatica, although applied works tend to be published in more specialized journals. The Winter Simulation Conference also publishes many relevant papers. Other than this, papers also published in the major conferences of the neural networks, fuzzy, and evolutionary computation communities. The annual IEEE symposium titled Approximate Dynamic Programming and Reinforcement Learning (ADPRL) and the biannual European Workshop on Reinforcement Learning (EWRL) are two regularly held meetings where RL researchers meet.