celeste/celeste_ai/celeste.py

@@ -28,9 +28,6 @@ class CelesteState(NamedTuple):
 	# Number of deaths since game start
 	deaths: int
 
-	# If an index is true, we got a strawberry on that stage.
-	berries: list[bool]
-
 	# Distance to next point
 	dist: float
 
@@ -226,7 +223,6 @@ class Celeste:
 				xvel			= float(self._internal_state["vx"]),
 				yvel			= float(self._internal_state["vy"]),
 				deaths			= int(self._internal_state["dc"]),
-				berries			= [x == "t" for x in self._internal_state["fr"][1:]],
 
 				dist			= self._dist,
 				next_point		= self._next_checkpoint_idx,

celeste/celeste_ai/train.py

@@ -15,10 +15,6 @@ if __name__ == "__main__":
 	# Where to read/write model data.
 	model_data_root = Path("model_data/current")
 
-	# Where PICO-8 saves screenshots.
-	# Probably your desktop.
-	screenshot_source = Path("/home/mark/Desktop")
-
 	model_save_path		= model_data_root / "model.torch"
 	model_archive_dir	= model_data_root / "model_archive"
 	model_train_log		= model_data_root / "train_log"
@@ -29,7 +25,7 @@ if __name__ == "__main__":
 
 
 	# Remove old screenshots
-	shots = screenshot_source.glob("hackcel_*.png")
+	shots = Path("/home/mark/Desktop").glob("hackcel_*.png")
 	for s in shots:
 		s.unlink()
 
@@ -284,6 +280,8 @@ def optimize_model():
 def on_state_before(celeste):
 	global steps_done
 
+	# Conversion to pytorch
+
 	state = celeste.state
 
 	pt_state = torch.tensor(
@@ -349,7 +347,7 @@ def on_state_after(celeste, before_out):
 		pt_next_state = None
 		reward = 0
 
-	# Reward for finishing a stage
+	# Reward for finishing stage
 	elif next_state.stage >= 1:
 		finished_stage = True
 		reward = next_state.next_point - state.next_point
@@ -372,7 +370,6 @@ def on_state_after(celeste, before_out):
 		if state.next_point == next_state.next_point:
 			reward = 0
 		else:
-			print(f"Got point {state.next_point}")
 			# Reward for reaching a point
 			reward = next_state.next_point - state.next_point
 
@@ -380,14 +377,6 @@ def on_state_after(celeste, before_out):
 			for i in range(state.next_point, state.next_point + reward):
 				point_counter[i] += 1
 
-		# Strawberry reward
-		if next_state.berries[state.stage] and not state.berries[state.stage]:
-			print(f"Got stage {state.stage} bonus")
-			reward += 1
-
-
-
-
 	reward = reward * 10
 	pt_reward = torch.tensor([reward], device = compute_device)
 
@@ -457,7 +446,7 @@ def on_state_after(celeste, before_out):
 
 
 		# Clean up screenshots
-		shots = screenshot_source.glob("hackcel_*.png")
+		shots = Path("/home/mark/Desktop").glob("hackcel_*.png")
 
 		target = screenshot_dir / Path(f"{episode_number}")
 		target.mkdir(parents = True)

celeste/resources/carts/hackcel.p8

@@ -1275,25 +1275,11 @@ function _update()
 	hack_has_sent_first_message = true
 	out_string = "dc:" .. tostr(deaths) .. ";"
 
-	-- Dash status
 	if hack_can_dash then
 		out_string = out_string .. "ds:t;"
 	else
 		out_string = out_string .. "ds:f;"
 	end
-	
-	
-	-- Fruit status
-	out_string = out_string .. "fr:"
-	for i = 0,29 do
-		if got_fruit[i] then
-			out_string = out_string .. "t"
-		else
-			out_string = out_string .. "f"
-		end
-	end
-	out_string = out_string .. ";"
-
 
 	for k, v in pairs(hack_player_state) do
 		out_string = out_string .. k ..":" .. v .. ";"