MirrorDescentActorCritic.jl

Tabular, Linear, and Deep Actor-Critic algorithms implemented with Lux.jl.

Installation

First, you'll need to get the ExtendedDistributions.jl package:

git clone git@github.com:samuelfneumann/ExtendedDistributions.jl.git
cd ExtendedDistributions
julia --project

then:

julia> ]instantiate

Then, clone this repo and jump into a Julia project

git clone git@github.com:samuelfneumann/MirrorDescentActorCritic.jl.git
cd MirrorDescentActorCritic.jl
julia --project

then

julia> ]instantiate

Algorithms

This codebase separates algorithms into actor and critic updates. Each algorithm is composed of each of these components, allowing different actor and critic updates to be composed easily to create new algorithms/agents. We provide implementations for the following standard actor updates, as well as their functional mirror descent counterparts:

- `CCEM`: Conditional Cross-Entropy Optimization [Link](https://arxiv.org/abs/1810.09103)
- `RKL`: Reverse KL to the Boltzmann (SAC-like) [Link](https://www.jmlr.org/papers/volume23/21-054/21-054.pdf)
- `FKL`: Forward KL to the Boltzmann [Link](https://www.jmlr.org/papers/volume23/21-054/21-054.pdf)
- `MPO`: Maximum A-Posteriori (MPO-like) [Link](https://www.jmlr.org/papers/volume23/21-054/21-054.pdf)
- `DualAveragingMPI`: Dual Averaging MPI [Link](https://arxiv.org/abs/2003.14089)
- `REINFORCE`: REINFORCE-like update
- `PPO`: PPO-like update

The following critic updates are implemented: - Sarsa

Critic updates can take a BellmanRegulariser, which alters the update to use a regularised Bellman equation. The following regularisers are implemented: - KLBellmanRegulariser: KL regularization to the previous policy - EntropyBellmanRegulariser: Entropy regularizer/soft value functions

Policies

This package has three kinds of continuous-action policies: BoundedPolicy, UnBoundedPolicy, and TruncatedPolicy.

To create a policy, you simply provide a policy struct with a distribution from the ExtendedDistributions.jl package. You should ensure that you provide bounded distributions to BoundedPolicy types and unbounded or half-bounded distributions to UnBoundedPolicy types.

Convenience constructors are provided for the following policies:

Distribution	Policy Type	RSample Supported
Kumaraswamy	BoundedPolicy	yes (quantile method)
Beta	BoundedPolicy	no
ArctanhNormal	BoundedPolicy	yes
LogitNormal	BoundedPolicy	yes
Normal	UnBoundedPolicy	yes
Laplace	UnBoundedPolicy	yes

Scheduling Experiments with SLURM

This codebase has been set up to easily schedule experiments with slurm.

There are two main directories for this, config and parallel. The config directory hold subdirectories of toml config files. The parallel directory holds subdirectories of sh jobscript files, which are scheduled with slurm through sbatch.

Each of these directories should be mirror images of each other. That is, if a file exists at config/X/Y/Z.toml, then a corresponding file should exists in parallel/X/Y/Z.sh. Here, config/X/Y/Z.toml outlines the experiment to run, including hyperparameter sweeps. Hyperparameters are swept using Reproduce.jl. The paralell/X/Y/Z.sh talks to slurm and tells it what to in order to run the experiment outlined by the corresponding config file.

To run an experiment, first create the config/X/Y/Z.toml file. A template is provided at config/template.toml. Then, create the corresponding parallel/X/Y/Z.sh file. A template is provided at parallel/template.toml. You'll need to fill in the preamble to tell slurm what resources you need. Then, fill in the lines marked with TODO. Finally, schedule the experiment using sbatch parallel/X/Y/Z.sh. Alternatively, you can use the scripts/seqbatch.sh file to automatically schedule sequential slurm jobs with dependencies. Note that if your experiment does not finish and times out, then rescheduling the experiment will continue from where the experiment left off.

Citing

If you find our work useful, please cite our preprint paper:

TODO