ReinforcementLearningCore.jl

Define how to select an action based on action values.

A hook is called at different stage during a run to allow users to inject customized runtime logic. By default, an AbstractHook will do nothing. One can customize the behavior by implementing the following methods:

By convention, the Base.getindex(h::YourHook) is implemented to extract the metrics we are interested in. Users can compose different AbstractHooks with +.

Abstract type for a learner.

The actor part must return logits (Do not use softmax in the last layer!), and the critic part must return a state value.

A wrapper of an AbstractPolicy. Generally speaking, it does nothing but to update the trajectory and policy appropriately in different stages. Agent is a Callable and its call method accepts varargs and keyword arguments to be passed to the policy.

Similar to StepsPerEpisode, but is specific to environments which return a Vector of rewards (a typical case with MultiThreadEnv).

CategoricalNetwork wraps a model (typically a neural network) that takes a state input and outputs logits for a categorical distribution. The optional argument mask must be an Array of Bool with the same size as state expect for the first dimension that must have the length of the action vector. Actions mapped to false by mask have a logit equal to -Inf and/or a zero-probability of being sampled.

Only applies if is_sampling is true and will return z, logits.

If is_sampling = false, returns only the logits obtained by a simple forward pass into model.

Sample action_samples actions from each state. Returns a 3D tensor with dimensions (action_size x action_samples x batchsize). Always returns the logits of each action along in a tensor with the same dimensions. The optional argument mask must be an Array of Bool with the same size as state expect for the first dimension that must have the length of the action vector. Actions mapped to false by mask have a logit equal to -Inf and/or a zero-probability of being sampled.

Returns μ and Σ when called where μ is the mean and Σ is a covariance matrix. Unlike GaussianNetwork, the output is 3-dimensional. μ has dimensions (action_size x 1 x batchsize) and Σ has dimensions (action_size x action_size x batchsize). The Σ head of the CovGaussianNetwork should not directly return a square matrix but a vector of length action_size x (action_size + 1) ÷ 2. This vector will contain elements of the uppertriangular cholesky decomposition of the covariance matrix, which is then reconstructed from it. Sample from MvNormal.(μ, Σ).

Return the logpdf of the model sampling action when in state. State must be a 3D tensor with dimensions (state_size x 1 x batchsize). Multiple actions may be taken per state, action must have dimensions (action_size x action_samples_per_state x batchsize). Returns a 3D tensor with dimensions (1 x action_samples_per_state x batchsize).

If given 2D matrices as input, will return a 2D matrix of logpdf. States and actions are paired column-wise, one action per state.

Sample action_samples actions per state in state and return the actions, logpdf(actions). This function is compatible with a multidimensional action space. The outputs are 3D tensors with dimensions (action_size x action_samples x batchsize) and (1 x action_samples x batchsize) for actions and logdpf respectively.

This function is compatible with a multidimensional action space. To work with covariance matrices, the outputs are 3D tensors. If sampling, return an actions tensor with dimensions (action_size x action_samples x batchsize) and a logp_π tensor with dimensions (1 x action_samples x batchsize). If not sampling, returns μ with dimensions (action_size x 1 x batchsize) and L, the lower triangular of the cholesky decomposition of the covariance matrix, with dimensions (action_size x action_size x batchsize) The covariance matrices can be retrieved with Σ = stack(map(l -> l*l', eachslice(L, dims=3)); dims=3)

Given a Matrix of states, will return actions, μ and logpdf in matrix format. The batch of Σ remains a 3D tensor.

CurrentPlayerIterator`is an iterator that iterates over the players in the environment, returning thecurrentplayerfor each iteration. This is only necessary for`MultiAgent`environments. After each iteration,`RLBase.next*player!`is called to advance the`current*player. As long asRLBase.nextplayer!is defined for the environment, this iterator will work correctly in theBase.run`` function.

Execute f(t, agent, env) every n episode. t is a counter of episodes.

Execute f(t, agent, env) every n step. t is a counter of steps.

Call the lambda function f at the end of an Experiment.

Dueling network automatically produces separate estimates of the state value function network and advantage function network. The expected output size of val is 1, and adv is the size of the action space.

Nothing but a placeholder.

Two kinds of epsilon-decreasing strategy are implemented here (linear and exp).

Keywords

Example

A struct to hold the information of an experiment. It is used to run an experiment with the given policy, environment, stop condition and hook.

Wraps a Flux trainable model and implements the RLBase.optimise!(::FluxApproximator, ::Gradient) interface. See the RLCore documentation for more information on proper usage.

Constructs an FluxApproximator object for reinforcement learning.

Arguments

Returns

An FluxApproximator object.

Sample action_samples actions from each state. Returns a 3D tensor with dimensions (action_size x action_samples x batchsize). state must be 3D tensor with dimensions (state_size x 1 x batchsize). Always returns the logpdf of each action along.

This function is compatible with a multidimensional action space.

MultiAgentHook is a hook struct that contains <:AbstractHook structs indexed by the player’s symbol.

MultiAgentPolicy is a policy struct that contains <:AbstractPolicy structs indexed by the player’s symbol.

OfflineAgent is an AbstractAgent that, unlike the usual online Agent, does not interact with the environment during training in order to collect data. Just like Agent, it contains an AbstractPolicy to be trained an a Trajectory that contains the training data. The difference being that the trajectory is filled prior to training and is not updated. An OfflineBehavior can optionally be provided to provide an second "behavior agent" that will generate the training data at the PreExperimentStage. Does nothing by default.

Used to provide an OfflineAgent with a "behavior agent" that will generate the training data at the PreExperimentStage. If agent is nothing (by default), does nothing. The trajectory of agent should be the same as that of the parent OfflineAgent. steps is the number of data elements to generate, defaults to the capacity of the trajectory. reset_condition is the episode reset condition for the data generation (defaults to ResetIfEnvTerminated()).

The behavior agent will interact with the main environment of the experiment to generate the data.

This function accepts state and action, and then outputs actions after disturbance.

A NamedTuple that maps players to their respective values.

Stage that is executed after the Agent acts.

Stage that is executed after the Episode is over.

Stage that is executed after the Experiment is over.

Stage that is executed before the Agent acts.

Stage that is executed before the Episode starts.

Stage that is executed before the Experiment starts.

Wraps a learner and an explorer. The learner is a struct that should predict the Q-value of each legal action of an environment at its current state. It is typically a table or a neural network. QBasedPolicy can be queried for an action with RLBase.plan!, the explorer will affect the action selection accordingly.

If action_space is nothing, then it will use the legal_action_space at runtime to randomly select an action. Otherwise, a random element within action_space is selected.

A reset condition that resets the environment after n steps.

ResetIfEnvTerminated()

A reset condition that resets the environment if is_terminated(env) is true.

Store each reward of each step in every episode in the field of rewards.

Like GaussianNetwork but with a differentiable reparameterization trick. Mainly used for SAC. Returns μ and σ when called. Create a distribution to sample from using Normal.(μ, σ). min_σ and max_σ are used to clip the output from σ. pre is a shared body before the two heads of the NN. σ should be > 0. You may enforce this using a softplus output activation. Actions are squashed by a tanh and a correction is applied to the logpdf.

Sample action_samples actions from each state. Returns a 3D tensor with dimensions (action_size x action_samples x batchsize). state must be 3D tensor with dimensions (state_size x 1 x batchsize). Always returns the logpdf of each action along.

This function is compatible with a multidimensional action space.

Use a pre-initialized CircularArrayBuffer to store the latest several states specified by d. Before processing any observation, the buffer is filled with `zero{T} by default.

Store steps of each episode in the field of steps.

Return true after being called episode. If is_show_progress is true, the ProgressMeter will be used to show progress.

StopAfterNSeconds

parameter:

stop training after N seconds

Return true after being called step times.

StopAfterNoImprovement()

Stop training when a monitored metric has stopped improving.

Parameters:

fn: a closure, return a scalar value, which indicates the performance of the policy (the higher the better) e.g.

patience: Number of epochs with no improvement after which training will be stopped.

δ: Minimum change in the monitored quantity to qualify as an improvement, i.e. an absolute change of less than min_delta, will count as no improvement.

Return true after the monitored metric has stopped improving.

The result of stop_conditions is reduced by any.

Return true if the environment is terminated.

Create a stop signal initialized with a value of false. You can manually set it to true by s[] = true to stop the running loop at any time.

Use temporal-difference method to estimate state value or state-action value.

Fields

For table of 1-d, it will serve as a state value approximator. For table of 2-d, it will serve as a state-action value approximator.

Create a TabularQApproximator with n_state states and n_action actions.

Wraps an FluxApproximator to hold a target network that is updated towards the model of the approximator.

The two common usages of TargetNetwork are

Implements the RLBase.optimise!(::TargetNetwork, ::Gradient) interface to update the model with the gradient and the target with weights replacement or Polyak averaging.

Note to developers: model(::TargetNetwork) will return the trainable Flux model and target(::TargetNetwork) returns the target model and target(::FluxApproximator) returns the non-trainable Flux model. See the RLCore documentation.

Constructs a target network for reinforcement learning.

Arguments

Returns

A TargetNetwork object.

Store time cost in seconds of the latest max_steps in the times field.

Store the total reward of each episode in the field of rewards. If is_display_on_exit is set to true, a unicode plot will be shown at the PostExperimentStage.

Arguments

is_normalized is used to indicate if the fed action values are already normalized to have a sum of 1.0.

See also: WeightedSoftmaxExplorer

See also: WeightedExplorer

When pushing a StackFrames into a CircularArrayBuffer of the same dimension, only the latest frame is pushed. If the StackFrames is one dimension lower, then it is treated as a general AbstractArray and is pushed in as a frame.

This run function dispatches games using MultiAgentPolicy and MultiAgentHook to the appropriate run function based on the Sequential or Simultaneous trait of the environment.

This run function handles MultiAgent games with the Sequential trait. It iterates over the current_player for each turn in the environment, and runs the full run loop, like in the SingleAgent case. If the stop_condition is met, the function breaks out of the loop and calls optimise! on the policy again. Finally, it calls optimise! on the policy one last time and returns the MultiAgentHook.

This run function handles MultiAgent games with the Simultaneous trait. It iterates over the players in the environment, and for each player, it selects the appropriate policy from the MultiAgentPolicy. All agent actions are collected before the environment is updated. After each player has taken an action, it calls optimise! on the policy. If the stop_condition is met, the function breaks out of the loop and calls optimise! on the policy again. Finally, it calls optimise! on the policy one last time and returns the MultiAgentHook.

Apply inner explorer to each column of values.

Similar to prob(p::AbstractExplorer, x), but here only the masked elements are considered.

Get the action distribution given action values.

Return the probability of selecting each action given the estimated values of each action.

assuming rewards and new_rewards are Vector

assuming rewards and advantages are Vector

Update the Q-value of the given state-action pair.

Inject some customized checkings here by overwriting this function

Return the position in a cholesky_vec (of length da) of the element of the lower triangular matrix at coordinates (i,j).

For example if cholesky_vec = [1,2,3,4,5,6], the corresponding lower triangular matrix is

and cholesky_matrix_to_vector_index(3, 2) == 5

GPU differentiable implementation of the kl_divergence between two MultiVariate Gaussian distributions with mean vectors μ1, μ2 respectively and diagonal standard deviations σ1, σ2. Arguments must be Vectors or arrays of column vectors.

GPU compatible and automatically differentiable version for the logpdf function of normal distributions with diagonal covariance. Adding an epsilon value to guarantee numeric stability if sigma is exactly zero (e.g. if relu is used in output layer). Accepts arguments of the same shape: vectors, matrices or 3D array (with dimension 2 of size 1).

Calculate the gain started from the current step with discount rate of γ. rewards can be a matrix.

Keyword arguments

Example

Merge the last two dimension.

Example

Calculate the generalized advantage estimate started from the current step with discount rate of γ and a lambda for GAE-Lambda of 'λ'. rewards and 'values' can be a matrix.

Keyword arguments

Example

Log-determinant of the Positive-Semi-Definite matrix A = L*U (cholesky lower and upper triangulars), given L or U. Has a sign uncertainty for non PSD matrices.

GPU differentiable implementation of the kl_divergence between two MultiVariate Gaussian distributions with mean vectors μ1, μ2 respectively and with cholesky decomposition of covariance matrices L1, L2.

GPU compatible and automatically differentiable version for the logpdf function of multivariate normal distributions. Takes as inputs mu the mean vector, L the lower triangular matrix of the cholesky decomposition of the covariance matrix, and x a matrix of samples where each column is a sample. Return a Vector containing the logpdf of each column of x for the MvNormal parametrized by μ and Σ = L*L'.

Batch version that takes 3D tensors as input where each slice along the 3rd dimension is a batch sample. μ is a (actionsize x 1 x batchsize) matrix, L is a (actionsize x actionsize x batchsize), x is a (actionsize x actionsamples x batchsize). Return a 3D matrix of size (1 x actionsamples x batchsize).

GPU differentiable implementation of the kl_divergence between two univariate Gaussian distributions with means μ1, μ2 and standard deviations σ1, σ2 respectively.

GPU automatic differentiable version for the logpdf function of a univariate normal distribution. Adding an epsilon value to guarantee numeric stability if sigma is exactly zero (e.g. if relu is used in output layer).

Transform a vector containing the non-zero elements of a lower triangular da x da matrix into that matrix.