diff --git a/marl_factory_grid/configs/default_config.yaml b/marl_factory_grid/configs/default_config.yaml index d3015c9..fe89597 100644 --- a/marl_factory_grid/configs/default_config.yaml +++ b/marl_factory_grid/configs/default_config.yaml @@ -22,13 +22,6 @@ Agents: - Inventory - DropOffLocations - Maintainers - # This is special for agents, as each one is differten and can act as an adversary e.g. - Positions: - - (16, 7) - - (16, 6) - - (16, 3) - - (16, 4) - - (16, 5) Entities: Batteries: initial_charge: 0.8 diff --git a/marl_factory_grid/environment/factory.py b/marl_factory_grid/environment/factory.py index d339403..b0a0372 100644 --- a/marl_factory_grid/environment/factory.py +++ b/marl_factory_grid/environment/factory.py @@ -91,20 +91,26 @@ class Factory(gym.Env): return self.state.entities[item] def reset(self) -> (dict, dict): + + # Reset information the state holds + self.state.reset() + + # Reset Information the GlobalEntity collection holds. self.state.entities.reset() # All is set up, trigger entity spawn with variable pos self.state.rules.do_all_reset(self.state) # Build initial observations for all agents - return self.obs_builder.refresh_and_build_for_all(self.state) + self.obs_builder.reset(self.state) + return self.obs_builder.build_for_all(self.state) def manual_step_init(self) -> List[Result]: self.state.curr_step += 1 # Main Agent Step pre_step_result = self.state.rules.tick_pre_step_all(self) - self.obs_builder.reset_struc_obs_block(self.state) + self.obs_builder.reset(self.state) return pre_step_result def manual_get_named_agent_obs(self, agent_name: str) -> (List[str], np.ndarray): @@ -158,7 +164,7 @@ class Factory(gym.Env): info.update(step_reward=sum(reward), step=self.state.curr_step) - obs = self.obs_builder.refresh_and_build_for_all(self.state) + obs = self.obs_builder.build_for_all(self.state) return None, [x for x in obs.values()], reward, done, info def summarize_step_results(self, tick_results: list, done_check_results: list) -> (int, dict, bool): diff --git a/marl_factory_grid/modules/destinations/rules.py b/marl_factory_grid/modules/destinations/rules.py index ef004c3..39d6714 100644 --- a/marl_factory_grid/modules/destinations/rules.py +++ b/marl_factory_grid/modules/destinations/rules.py @@ -106,7 +106,7 @@ class SpawnDestinationsPerAgent(Rule): super(Rule, self).__init__() self.per_agent_positions = {key: [ast.literal_eval(x) for x in val] for key, val in coords_or_quantity.items()} - def on_reset(self, state, lvl_map): + def on_reset(self, state): for (agent_name, position_list) in self.per_agent_positions.items(): agent = h.get_first(state[c.AGENT], lambda x: agent_name in x.name) assert agent diff --git a/marl_factory_grid/utils/observation_builder.py b/marl_factory_grid/utils/observation_builder.py index 55d6ec0..35497bf 100644 --- a/marl_factory_grid/utils/observation_builder.py +++ b/marl_factory_grid/utils/observation_builder.py @@ -24,11 +24,7 @@ class OBSBuilder(object): return 0 def __init__(self, level_shape: np.size, state: Gamestate, pomdp_r: int): - self._curr_env_step = None self.all_obs = dict() - self.light_blockers = defaultdict(lambda: False) - self.positional = defaultdict(lambda: False) - self.non_positional = defaultdict(lambda: False) self.ray_caster = dict() self.level_shape = level_shape @@ -37,13 +33,15 @@ class OBSBuilder(object): self.size = np.prod(self.obs_shape) self.obs_layers = dict() - - self.reset_struc_obs_block(state) self.curr_lightmaps = dict() + self._floortiles = defaultdict(list, {pos: [Floor(*pos)] for pos in state.entities.floorlist}) - def reset_struc_obs_block(self, state): - self._curr_env_step = state.curr_step + self.reset(state) + + def reset(self, state): + # Reset temporary information + self.curr_lightmaps = dict() # Construct an empty obs (array) for possible placeholders self.all_obs[c.PLACEHOLDER] = np.full(self.obs_shape, 0, dtype=float) # Fill the all_obs-dict with all available entities @@ -52,7 +50,8 @@ class OBSBuilder(object): def observation_space(self, state): from gymnasium.spaces import Tuple, Box - obsn = self.refresh_and_build_for_all(state) + self.reset(state) + obsn = self.build_for_all(state) if len(state[c.AGENT]) == 1: space = Box(low=0, high=1, shape=next(x for x in obsn.values()).shape, dtype=np.float32) else: @@ -60,14 +59,13 @@ class OBSBuilder(object): return space def named_observation_space(self, state): - return self.refresh_and_build_for_all(state) + self.reset(state) + return self.build_for_all(state) - def refresh_and_build_for_all(self, state) -> (dict, dict): - self.reset_struc_obs_block(state) + def build_for_all(self, state) -> (dict, dict): return {agent.name: self.build_for_agent(agent, state)[0] for agent in state[c.AGENT]} - def refresh_and_build_named_for_all(self, state) -> Dict[str, Dict[str, np.ndarray]]: - self.reset_struc_obs_block(state) + def build_named_for_all(self, state) -> Dict[str, Dict[str, np.ndarray]]: named_obs_dict = {} for agent in state[c.AGENT]: obs, names = self.build_for_agent(agent, state) @@ -85,9 +83,6 @@ class OBSBuilder(object): pass def build_for_agent(self, agent, state) -> (List[str], np.ndarray): - assert self._curr_env_step == state.curr_step, ( - "The observation objekt has not been reset this state! Call 'reset_struc_obs_block(state)'" - ) try: agent_want_obs = self.obs_layers[agent.name] except KeyError: @@ -166,7 +161,8 @@ class OBSBuilder(object): raise ValueError(f'Max(obs.size) for {e.name}: {obs[idx].size}, but was: {len(v)}.') if self.pomdp_r: try: - light_map = np.zeros(self.obs_shape) + light_map = self.curr_lightmaps.get(agent.name, np.zeros(self.obs_shape)) + light_map[:] = 0.0 visible_floor = self.ray_caster[agent.name].visible_entities(self._floortiles, reset_cache=False) for f in set(visible_floor): diff --git a/marl_factory_grid/utils/plotting/plotting_utils.py b/marl_factory_grid/utils/plotting/plotting_utils.py index 17bb7ff..2ae61f0 100644 --- a/marl_factory_grid/utils/plotting/plotting_utils.py +++ b/marl_factory_grid/utils/plotting/plotting_utils.py @@ -49,7 +49,7 @@ def prepare_plt(df, hue, style, hue_order): plt.close('all') sns.set(rc={'text.usetex': False}, style='whitegrid') lineplot = sns.lineplot(data=df, x='Episode', y='Score', hue=hue, style=style, - ci=95, palette=PALETTE, hue_order=hue_order, ) + errorbar=('ci', 95), palette=PALETTE, hue_order=hue_order, ) plt.legend(bbox_to_anchor=(1.02, 1), loc='upper left', borderaxespad=0) plt.tight_layout() # lineplot.set_title(f'{sorted(list(df["Measurement"].unique()))}') diff --git a/marl_factory_grid/utils/states.py b/marl_factory_grid/utils/states.py index d6db02a..2a363bf 100644 --- a/marl_factory_grid/utils/states.py +++ b/marl_factory_grid/utils/states.py @@ -86,6 +86,10 @@ class Gamestate(object): self.rules = StepRules(*rules) self._floortile_graph = None + def reset(self): + self.curr_step = 0 + self.curr_actions = None + def __getitem__(self, item): return self.entities[item] diff --git a/random_testrun.py b/random_testrun.py index ef8df08..3f6c56f 100644 --- a/random_testrun.py +++ b/random_testrun.py @@ -12,7 +12,7 @@ from marl_factory_grid.utils.tools import ConfigExplainer if __name__ == '__main__': # Render at each step? - render = False + render = True # Reveal all possible Modules (Entities, Rules, Agents[Actions, Observations], etc.) explain_config = False # Collect statistics?