A PyTorch library for building deep reinforcement learning agents.
Version 0.9.1. Includes the following updates:
Full Changelog: https://github.com/cpnota/autonomous-learning-library/compare/v0.8.0...v0.9.1
Publish workflow testing
This release includes several enhancements:
RuntimeError
if clip_grad
is enabled and norm is non-finite (#255)log_prob
in soft policy (#256)A few minor under-the-hood tweaks and fixes:
VectorEnvironment
class and refactored the way parallel environments work #239parallel_test_agent
method to ParallelPreset
#240n_envs=1
on Atari presets used Body
instead of ParallelBody
#241DeepmindAtariBody
to only use FrameStackBody
if frame_stack > 1
#245test_exploration
was not being respected by GreedyPolicy
#246store_device
for prioritized replay #249This release contains several new features, refactors, and bugfixes.
Identity
feature network #202Agent
, ParallelAgent
, and Multiagent
#221Environment
imports #236This release contains some under-the-hood enhancements and bugfixes, most notably, a refactoring of the State
class. State
now supports adding arbitrary key/value pairs, allowing for more complex state spaces. Additionally, a StateArray
class was added that automatically handles stacking/slicing states in various ways which making handling batches of data, multiple timesteps, and many other aspects easier. Here's a full list of the changes:
State
and added StateArray
object ( #160 and #167)FireReset
wrapper was being applied to games with no Fire action, causing them to not run. Thanks @andrewsmike for reporting and fixing the bug! (#168)This release contains a hotfix, #155 , which improves the performance of the PPO continuous preset.
Just some minor bug fixes and documentation improvements:
The previous release was missing the changes from #132 with correct the computation of the frames per second for parallel envs.