Same as battle but with fewer agents arrayed in a larger space with obstacles.

A small-scale team battle. Agents have to figure out the optimal way to coordinate their small team in a large space and maneuver around obstacles in order to defeat the opponent team.

Like all MAgent environments, agents can either move or attack each turn. An attack against another agent on their own team will not be registered.

Action options: [do_nothing, move_12, attack_8]

Reward is given as:

• 5 reward for killing an opponent
• -0.005 reward every step (step_reward option)
• -0.1 reward for attacking (attack_penalty option)
• 0.2 reward for attacking an opponent (attack_opponent_reward option)
• -0.1 reward for dying (dead_penalty option)

If multiple options apply, rewards are added together.

Observation space: [empty, obstacle, red, blue, minimap_red, minimap_blue, binary_agent_id(10), one_hot_action, last_reward, agent_position]

Map size: 80x80

Arguments

battle_v1.env(step_reward-0.005, dead_penalty=-0.1, attack_penalty=-0.1, attack_opponent_reward=0.2, max_frames=1000)

step_reward: reward added unconditionally

dead_penalty: reward added when killed

attack_penalty: reward added for attacking

attack_opponent_reward: Reward added for attacking an opponent

max_frames: number of frames (a step for each agent) until game terminates