您最多选择25个主题
主题必须以中文或者字母或数字开头,可以包含连字符 (-),并且长度不得超过35个字符
目录树:
12753429
/main
/develop-generalizationTraining-TrainerController
/tag-0.2.0
/tag-0.2.1
/tag-0.2.1a
/tag-0.2.1c
/tag-0.2.1d
/hotfix-v0.9.2a
/develop-gpu-test
/0.10.1
/develop-pyinstaller
/develop-horovod
/PhysXArticulations20201
/importdocfix
/develop-resizetexture
/hh-develop-walljump_bugfixes
/develop-walljump-fix-sac
/hh-develop-walljump_rnd
/tag-0.11.0.dev0
/develop-pytorch
/tag-0.11.0.dev2
/develop-newnormalization
/tag-0.11.0.dev3
/develop
/release-0.12.0
/tag-0.12.0-dev
/tag-0.12.0.dev0
/tag-0.12.1
/2D-explorations
/asymm-envs
/tag-0.12.1.dev0
/2D-exploration-raycast
/tag-0.12.1.dev1
/release-0.13.0
/release-0.13.1
/plugin-proof-of-concept
/release-0.14.0
/hotfix-bump-version-master
/soccer-fives
/release-0.14.1
/bug-failed-api-check
/test-recurrent-gail
/hh-add-icons
/release-0.15.0
/release-0.15.1
/hh-develop-all-posed-characters
/internal-policy-ghost
/distributed-training
/hh-develop-improve_tennis
/test-tf-ver
/release_1_branch
/tennis-time-horizon
/whitepaper-experiments
/r2v-yamato-linux
/docs-update
/release_2_branch
/exp-mede
/sensitivity
/release_2_verified_load_fix
/test-sampler
/release_2_verified
/hh-develop-ragdoll-testing
/origin-develop-taggedobservations
/MLA-1734-demo-provider
/sampler-refactor-copy
/PhysXArticulations20201Package
/tag-com.unity.ml-agents_1.0.8
/release_3_branch
/github-actions
/release_3_distributed
/fix-batch-tennis
/distributed-ppo-sac
/gridworld-custom-obs
/hw20-segmentation
/hh-develop-gamedev-demo
/active-variablespeed
/release_4_branch
/fix-env-step-loop
/release_5_branch
/fix-walker
/release_6_branch
/hh-32-observation-crawler
/trainer-plugin
/hh-develop-max-steps-demo-recorder
/hh-develop-loco-walker-variable-speed
/exp-0002
/experiment-less-max-step
/hh-develop-hallway-wall-mesh-fix
/release_7_branch
/exp-vince
/hh-develop-gridsensor-tests
/tag-release_8_test0
/tag-release_8_test1
/release_8_branch
/docfix-end-episode
/release_9_branch
/hybrid-action-rewardsignals
/MLA-462-yamato-win
/exp-alternate-atten
/hh-develop-fps_game_project
/fix-conflict-base-env
/release_10_branch
/exp-bullet-hell-trainer
/ai-summit-exp
/comms-grad
/walljump-pushblock
/goal-conditioning
/release_11_branch
/hh-develop-water-balloon-fight
/gc-hyper
/layernorm
/yamato-linux-debug-venv
/soccer-comms
/hh-develop-pushblockcollab
/release_12_branch
/fix-get-step-sp-curr
/continuous-comms
/no-comms
/hh-develop-zombiepushblock
/hypernetwork
/revert-4859-develop-update-readme
/sequencer-env-attention
/hh-develop-variableobs
/exp-tanh
/reward-dist
/exp-weight-decay
/exp-robot
/bullet-hell-barracuda-test-1.3.1
/release_13_branch
/release_14_branch
/exp-clipped-gaussian-entropy
/tic-tac-toe
/hh-develop-dodgeball
/repro-vis-obs-perf
/v2-staging-rebase
/release_15_branch
/release_15_removeendepisode
/release_16_branch
/release_16_fix_gridsensor
/ai-hw-2021
/check-for-ModelOverriders
/fix-grid-obs-shape-init
/fix-gym-needs-reset
/fix-resume-imi
/release_17_branch
/release_17_branch_gpu_test
/colab-links
/exp-continuous-div
/release_17_branch_gpu_2
/exp-diverse-behavior
/grid-onehot-extra-dim-empty
/2.0-verified
/faster-entropy-coeficient-convergence
/pre-r18-update-changelog
/release_18_branch
/main/tracking
/main/reward-providers
/main/project-upgrade
/main/limitation-docs
/develop/nomaxstep-test
/develop/tf2.0
/develop/tanhsquash
/develop/magic-string
/develop/trainerinterface
/develop/separatevalue
/develop/nopreviousactions
/develop/reenablerepeatactions
/develop/0memories
/develop/fixmemoryleak
/develop/reducewalljump
/develop/removeactionholder-onehot
/develop/canonicalize-quaternions
/develop/self-playassym
/develop/demo-load-seek
/develop/progress-bar
/develop/sac-apex
/develop/cubewars
/develop/add-fire
/develop/gym-wrapper
/develop/mm-docs-main-readme
/develop/mm-docs-overview
/develop/no-threading
/develop/dockerfile
/develop/model-store
/develop/checkout-conversion-rebase
/develop/model-transfer
/develop/bisim-review
/develop/taggedobservations
/develop/transfer-bisim
/develop/bisim-sac-transfer
/develop/basketball
/develop/torchmodules
/develop/fixmarkdown
/develop/shortenstrikervsgoalie
/develop/shortengoalie
/develop/torch-save-rp
/develop/torch-to-np
/develop/torch-omp-no-thread
/develop/actionmodel-csharp
/develop/torch-extra
/develop/restructure-torch-networks
/develop/jit
/develop/adjust-cpu-settings-experiment
/develop/torch-sac-threading
/develop/wb
/develop/amrl
/develop/memorydump
/develop/permutepytorch
/develop/sac-targetq
/develop/actions-out
/develop/reshapeonnxmemories
/develop/crawlergail
/develop/debugtorchfood
/develop/hybrid-actions
/develop/bullet-hell
/develop/action-spec-gym
/develop/battlefoodcollector
/develop/use-action-buffers
/develop/hardswish
/develop/leakyrelu
/develop/torch-clip-scale
/develop/contentropy
/develop/manch
/develop/torchcrawlerdebug
/develop/fix-nan
/develop/multitype-buffer
/develop/windows-delay
/develop/torch-tanh
/develop/gail-norm
/develop/multiprocess
/develop/unified-obs
/develop/rm-rf-new-models
/develop/skipcritic
/develop/centralizedcritic
/develop/dodgeball-tests
/develop/cc-teammanager
/develop/weight-decay
/develop/singular-embeddings
/develop/zombieteammanager
/develop/superpush
/develop/teammanager
/develop/zombie-exp
/develop/update-readme
/develop/readme-fix
/develop/coma-noact
/develop/coma-withq
/develop/coma2
/develop/action-slice
/develop/gru
/develop/critic-op-lstm-currentmem
/develop/decaygail
/develop/gail-srl-hack
/develop/rear-pad
/develop/mm-copyright-dates
/develop/dodgeball-raycasts
/develop/collab-envs-exp-ervin
/develop/pushcollabonly
/develop/sample-curation
/develop/soccer-groupman
/develop/input-actuator-tanks
/develop/validate-release-fix
/develop/new-console-log
/develop/lex-walker-model
/develop/lstm-burnin
/develop/grid-vaiable-names
/develop/fix-attn-embedding
/develop/api-documentation-update-some-fixes
/develop/update-grpc
/develop/grid-rootref-debug
/develop/pbcollab-rays
/develop/2.0-verified-pre
/develop/parameterizedenvs
/develop/custom-ray-sensor
/develop/mm-add-v2blog
/develop/custom-raycast
/develop/area-manager
/develop/remove-unecessary-lr
/develop/use-base-env-in-learn
/soccer-fives/multiagent
/develop/cubewars/splashdamage
/develop/add-fire/exp
/develop/add-fire/jit
/develop/add-fire/speedtest
/develop/add-fire/bc
/develop/add-fire/ckpt-2
/develop/add-fire/normalize-context
/develop/add-fire/components-dir
/develop/add-fire/halfentropy
/develop/add-fire/memoryclass
/develop/add-fire/categoricaldist
/develop/add-fire/mm
/develop/add-fire/sac-lst
/develop/add-fire/mm3
/develop/add-fire/continuous
/develop/add-fire/ghost
/develop/add-fire/policy-tests
/develop/add-fire/export-discrete
/develop/add-fire/test-simple-rl-fix-resnet
/develop/add-fire/remove-currdoc
/develop/add-fire/clean2
/develop/add-fire/doc-cleanups
/develop/add-fire/changelog
/develop/add-fire/mm2
/develop/model-transfer/add-physics
/develop/model-transfer/train
/develop/jit/experiments
/exp-vince/sep30-2020
/hh-develop-gridsensor-tests/static
/develop/hybrid-actions/distlist
/develop/bullet-hell/buffer
/goal-conditioning/new
/goal-conditioning/sensors-2
/goal-conditioning/sensors-3-pytest-fix
/goal-conditioning/grid-world
/soccer-comms/disc
/develop/centralizedcritic/counterfact
/develop/centralizedcritic/mm
/develop/centralizedcritic/nonego
/develop/zombieteammanager/disableagent
/develop/zombieteammanager/killfirst
/develop/superpush/int
/develop/superpush/branch-cleanup
/develop/teammanager/int
/develop/teammanager/cubewar-nocycle
/develop/teammanager/cubewars
/develop/superpush/int/hunter
/goal-conditioning/new/allo-crawler
/develop/coma2/clip
/develop/coma2/singlenetwork
/develop/coma2/samenet
/develop/coma2/fixgroup
/develop/coma2/samenet/sum
/hh-develop-dodgeball/goy-input
/develop/soccer-groupman/mod
/develop/soccer-groupman/mod/hunter
/develop/soccer-groupman/mod/hunter/cine
/ai-hw-2021/tensor-applier
0.1.1
0.1.2
0.10.0
0.10.1
0.11.0
0.11.0.dev0
0.11.0.dev1
0.11.0.dev2
0.11.0.dev3
0.12.0
0.12.0-dev
0.12.0.dev0
0.12.1
0.12.1.dev0
0.12.1.dev1
0.13.0
0.13.1
0.14.0
0.14.1
0.15.0
0.15.1
0.2.0
0.2.1
0.2.1a
0.2.1b
0.2.1c
0.2.1d
0.3.0
0.3.0a
0.3.0b
0.3.1
0.3.1a
0.3.1b
0.4.0
0.4.0a
0.4.0b
0.5.0
0.5.0a
0.6.0
0.6.0a
0.7.0
0.8.0
0.8.1
0.8.2
0.9.0
0.9.1
0.9.2
0.9.3
com.unity.ml-agents_1.0.0
com.unity.ml-agents_1.0.2
com.unity.ml-agents_1.0.3
com.unity.ml-agents_1.0.4
com.unity.ml-agents_1.0.5
com.unity.ml-agents_1.0.6
com.unity.ml-agents_1.0.7
com.unity.ml-agents_1.0.8
com.unity.ml-agents_1.1.0
com.unity.ml-agents_1.2.0
com.unity.ml-agents_1.3.0
com.unity.ml-agents_1.4.0
com.unity.ml-agents_1.5.0
com.unity.ml-agents_1.6.0
com.unity.ml-agents_1.7.0
com.unity.ml-agents_1.7.2
com.unity.ml-agents_1.8.0
com.unity.ml-agents_1.8.1
com.unity.ml-agents_1.9.0
com.unity.ml-agents_1.9.1
com.unity.ml-agents_2.0.0
com.unity.ml-agents_2.1.0
latest_release
python-packages_0.16.0
python-packages_0.16.1
python-packages_0.17.0
python-packages_0.18.0
python-packages_0.18.1
python-packages_0.19.0
python-packages_0.20.0
python-packages_0.21.0
python-packages_0.21.1
python-packages_0.22.0
python-packages_0.23.0
python-packages_0.24.0
python-packages_0.24.1
python-packages_0.25.0
python-packages_0.25.1
python-packages_0.26.0
python-packages_0.27.0
release_1
release_10
release_10_docs
release_11
release_11_docs
release_12
release_12_docs
release_13
release_13_docs
release_14
release_14_docs
release_15
release_15_docs
release_16
release_16_docs
release_17
release_17_docs
release_18
release_18_docs
release_1_docs
release_2
release_2_docs
release_2_verified_docs
release_3
release_3_docs
release_4
release_4_docs
release_5
release_5_docs
release_6
release_6_docs
release_7
release_7_docs
release_8
release_8_docs
release_8_test0
release_8_test1
release_9
release_9_docs
v0.1
1.2 KiB
1.2 KiB
Environment Design Best Practices
General
- It is often helpful to being with the simplest version of the problem, to ensure the agent can learn it. From there increase complexity over time.
- When possible, It is often helpful to ensure that you can complete the task by using a Player Brain to control the agent.
Rewards
- The magnitude of any given reward should typically not be greater than 1.0 in order to ensure a more stable learning process.
- Positive rewards are often more helpful to shaping the desired behavior of an agent than negative rewards.
- For locomotion tasks, a small positive reward (+0.1) for forward progress is typically used.
- If you want the agent the finish a task quickly, it is often helpful to provide a small penalty every step (-0.1).
States
- The magnitude of each state variable should be normalized to around 1.0.
- States should include all variables relevant to allowing the agent to take the optimally informed decision.
- Categorical state variables such as type of object (Sword, Shield, Bow) should be encoded in one-hot fashion (ie
3
->0, 0, 1
).
Actions
- When using continuous control, action values should be clipped to an appropriate range.