AltitudeReward Punish if current fighter doesn't satisfy some constraints. Typically negative. - Punishment of velocity when lower than s

EventDrivenReward Achieve reward when the following event happens: - Shot down by missile: -200 - Crash accidentally: -200 - Shoo

PostureReward = Orientation * Range - Orientation: Encourage pointing at enemy fighter, punish when is pointed at. - Range: Encourage get

Plain MLP

MissilePostureReward Use the velocity attenuation

Plain CNN

A linear layer followed by a Categorical distribution.

Q Network for continuous and discrete action space. Outputs the q value given global states and actions. Note that the name ContinuousQNet emphasi

ExtremeState End up the simulation if the aircraft is on an extreme state.

LowAltitude End up the simulation if altitude are too low.

Overload End up the simulation if acceleration are too high.

SingleCombatEnv is an one-to-one competitive environment.

The StarCraft II environment for decentralised multi-agent micromanagement scenarios.

Timeout Episode terminates if max_step steps have passed.

MLP Module to compute actions.

Dueling Q Network for discrete action space.

SafeReturn. End up the simulation if: - the current aircraft has been shot down. - all the enemy-aircrafts has been destroyed

ShootPenaltyReward when launching a missile, give -10 reward for penalty, to avoid launching all missiles at once

Stochastic policy model. Outputs actions given observations.

Normalize a vector of observations - across the first norm_axes dimensions

A class which wraps an instance of JSBSim and manages communication with it.

Base Task class. A class to subclass in order to create a task with its own observation variables, action variables, termination conditio

Uses cloudpickle to serialize contents (otherwise multiprocessing tries to use pickle)

Deterministic policy network for continuous action space.

A linear layer followed by a Diagonal Gaussian distribution.

Modify standard PyTorch Categorical.

Modify standard PyTorch Normal.

Measure the difference between the current heading and the target heading

Control target heading with discrete action space

A class to store both lag & extra properties initiated and used during lag simulation.

MultipleCombatEnv is an multi-player competitive environment.

Off-policy buffer that uses Environment-Provided (EP) state.

Off-policy buffer that uses Feature-Pruned (FP) state. When FP state is used, the critic takes different global state as input for different actor

On-policy buffer for actor data storage.

On-policy buffer for critic that uses Environment-Provided (EP) state.

On-policy buffer for critic that uses Feature-Pruned (FP) state. When FP state is used, the critic takes different global state as input for diffe

This task aims at training agent to dodge missile attacking

SingleControlEnv is an fly-control env for single agent with no enemy fighters.

Squashed Gaussian policy network for HASAC.

Stochastic policy model that only uses MLP network. Outputs actions given observations.

UnreachHeading [0, 1] End up the simulation if the aircraft didn't reach the target heading or attitude in limited time.

V Critic. Critic that learns a V-function.

V Network. Outputs value function predictions given global states.

A class wrapping the JSBSim flight dynamics module (FDM) for simulating aircraft as an RL environment conforming to the OpenAI Gym Env in

Base logger class. Used for logging information in the on-policy training pipeline.

Base RewardFunction class Reward-specific reset and get_reward methods are implemented in subclasses

Base TerminationCondition class Condition-specific get_termination method is implemented in subclasses