Guide to ML Agents 1.0

• Independent agents, but share the same Behavior

  • Many instances help to speed up training in parallel
  • Behaviors determines how agents makes decisions

• Agent Properties

  • Max Step: how many simulation steps before episode ends. Agents restarts every N steps.
  • Behaviors parameters
    • Vector Observation Space
    • Vector Action Space (Discrete/Continuous)

• Mean reward should increase during training

• Should end the training from Python (keyboard event for ctrl-c) to trigger saving the .nn model file

• TODO Add support for --resume to load up checkpoints

• To train behaviors we need to define 3 entities for a given environment

  • Observations
  • Actions
  • Reward signals

• By training, an agents learns a policy, which is an optional map between observations into actions

• Heuristic behaviors is maybe a good place to put our navmesh-based control for "classical" AI (eg if agent collides, turn 90deg and continue forward) Behaviors defined as hard-coded set of rules

• Behaviors are like a function: f(observations) = actions

• Goal of agent: discover a behavior (a Policy) that maximizes a reward

  • Rewards
    • Extrinsic
    • Intrinsic
      • GAIL (Generative Adversarial IL)
      • Behavior Cloning (BC, can be enabled on PPO or SAC)
        • Works better when there's demonstrations of almost all agent states
      • Curiosity

• RL algorithms

  • Proximal Policy Optimization
  • Soft Actor-Critic
    • Heaver/slower (0.1 sec per step)

• Imitation Learning: we can combine with RL to dramatically reduce the time the agent takes to solve the environment.

  • Recording demonstrations on editor or build, saving as assets

To summarize, we provide 3 training methods: BC, GAIL and RL (PPO or SAC) that can be used independently or together:

• BC can be used on its own or as a pre-training step before GAIL and/or RL
• GAIL can be used with or without extrinsic rewards
• RL can be used on its own (either PPO or SAC) or in conjunction with BC and/or GAIL.

Leveraging either BC or GAIL requires recording demonstrations to be provided as input to the training algorithms.

• Training with curriculum

  • We can add some incremental variations of parameters (eg wall height) to increase difficulty level to agents while training

• Environment Parameter Randomization

  • Training agents by varying environment parameters, making them more robust to avoid overfiting

• Model types

  • Vector Observations
    • Include a fully connected NN to learn from vector observations
      • Visual Observations
    • Include 3 CNN architectures to learn from multiple-cameras
      • Simple encoder with 2 conv layers
      • 3 conv layers proposed by Mnih et al.
      • 3 stacked layers of 2 resnets (IMPALA Resnet), much larter than other two
  • Memory-enhanced (LSTM)

Parameters: https://github.com/Unity-Technologies/ml-agents/blob/release_1/docs/Training-Configuration-File.md

PPO

Proximal Policy Optimization

• Tried to eliminate stepsize flaw on Plicy gradient methods by developing TRPO and ACER
  • ACER is far more complicated than PPO (replay buffer)

Important Parameters