Changeset 24962

Timestamp:

Feb 28, 2021, 1:16:32 PM (3 years ago)

Author:

wraitii

Message:

Add an 'Evaluate' RL endpoint, intended for dynamic game balancing

This adds a new endpoint to the RL interface for evaluating custom JavaScript. When combined with the ability to create arbitrary modifiers, this provides the required functionality for exploring quantitative game balancing.

Patch By: irishninja

Fixes #5981

Differential Revision: ​https://code.wildfiregames.com/D3479

Location:

ps/trunk/source

Files:

• rlinterface/RLInterface.cpp (modified) (7 diffs)
• rlinterface/RLInterface.h (modified) (8 diffs)
• tools/rlclient/python/tests/fastactions.js (added)
• tools/rlclient/python/tests/test_actions.py (modified) (1 diff)
• tools/rlclient/python/tests/test_evaluate.py (added)
• tools/rlclient/python/zero_ad/api.py (modified) (1 diff)
• tools/rlclient/python/zero_ad/environment.py (modified) (1 diff)

ps/trunk/source/rlinterface/RLInterface.cpp

@@ -63 +63 @@
     m_GameMessage = std::move(msg);
     m_MsgApplied.wait(msgLock, [this]() { return m_GameMessage.type == GameMessageType::None; });
-    return m_GameState;
+    return m_e;
 }
 
@@ -77 +77 @@
     m_ScenarioConfig = std::move(scenario);
     return SendGameMessage({ GameMessageType::Reset });
+
+
+
+
+
+
+
 }
 
@@ -193 +200 @@
                 stream << gameState.c_str();
         }
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
         else if (uri == "/templates")
         {
@@ -275 +306 @@
     if (m_NeedsGameState && isGameStarted)
     {
-        m_GameState = GetGameState();
+        m_e = GetGameState();
         m_MsgApplied.notify_one();
         m_MsgLock.unlock();
@@ -319 +350 @@
                 LDR_NonprogressiveLoad();
                 ENSURE(g_Game->ReallyStartGame() == PSRETURN_OK);
-                m_GameState = GetGameState();
+                m_e = GetGameState();
                 m_MsgApplied.notify_one();
                 m_MsgLock.unlock();
@@ -343 +374 @@
             if (!g_Game)
             {
-                m_GameState = EMPTY_STATE;
+                m_e = EMPTY_STATE;
                 m_MsgApplied.notify_one();
                 m_MsgLock.unlock();
@@ -369 +400 @@
                 g_Game->Update(deltaRealTime);
 
-            m_GameState = GetGameState();
+            m_e = GetGameState();
             m_MsgApplied.notify_one();
             m_MsgLock.unlock();
             break;
         }
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
         default:
         break;

ps/trunk/source/rlinterface/RLInterface.h

@@ -1 @@
-/* Copyright (C) 2020 Wildfire Games.
+/* Copyright (C) 202 Wildfire Games.
  * This file is part of 0 A.D.
  *
@@ -46 +46 @@
     Reset,
     Commands,
+
 };
 
@@ -72 +73 @@
  *  0. The game/main thread calls TryApplyMessage()
  *    - If no messages are pending, GOTO 0 (the simulation is not advanced).
- *  1. TryApplyMessage locks m_MsgLock, pulls the message, processes it, advances the simulation, and sets m_GameState.
+ *  1. TryApplyMessage locks m_MsgLock, pulls the message, processes it, advances the simulation, and sets m_e.
  *  2. TryApplyMessage notifies the RL thread that it can carry on and unlocks m_MsgLock. The main thread carries on frame rendering and goes back to 0.
- *  3. The RL thread locks m_MsgLock, reads m_GameState, unlocks m_MsgLock, and sends the gamestate as HTTP Response to the RL client.
+ *  3. The RL thread locks m_MsgLock, reads m_e, unlocks m_MsgLock, and sends the gamestate as HTTP Response to the RL client.
  *  4. The client processes the response and ultimately sends a new HTTP message to the RL Interface.
  *  5. The RL thread locks m_MsgLock, pushes the message, and starts waiting on the game/main thread to notify it (step 2).
@@ -87 +88 @@
     /**
      * Non-blocking call to process any pending messages from the RL client.
-     * Updates m_GameState to the gamestate after messages have been processed.
+     * Updates m_e to the gamestate after messages have been processed.
      */
     void TryApplyMessage();
@@ -106 +107 @@
      */
     std::string Reset(ScenarioConfig&& scenario);
+
+
+
+
+
+
 
     /**
@@ -132 +139 @@
     /**
      * Process any pending messages from the RL client.
-     * Updates m_GameState to the gamestate after messages have been processed.
+     * Updates m_e to the gamestate after messages have been processed.
      */
     void ApplyMessage(const GameMessage& msg);
@@ -145 +152 @@
     GameMessage m_GameMessage;
     ScenarioConfig m_ScenarioConfig;
-    std::string m_GameState;
+    std::string m_e;
     bool m_NeedsGameState = false;
 
@@ -151 +158 @@
     std::mutex m_MsgLock;
     std::condition_variable m_MsgApplied;
+
 };
 

ps/trunk/source/tools/rlclient/python/tests/test_actions.py

@@ -79 +79 @@
 def test_attack():
     state = game.reset(config)
-    units = state.units(owner=1, type='cavalry')
+    unit
     target = state.units(owner=2, type='female_citizen')[0]
-    initial_health = target.health()
+    initial_health_target = target.health()
+    initial_health_unit = unit.health()
 
     state = game.step([zero_ad.actions.reveal_map()])
 
-    attack = zero_ad.actions.attack(units, target)
+    attack = zero_ad.actions.attack(, target)
     state = game.step([attack])
-    while state.unit(target.id()).health() >= initial_health:
+    while (state.unit(target.id()).health() >= initial_health_target
+        ) and (state.unit(unit.id()).health() >= initial_health_unit):
         state = game.step()
 

ps/trunk/source/tools/rlclient/python/zero_ad/api.py

@@ -28 +28 @@
         response = self.post('templates', post_data)
         return zip(names, response.decode().split('\n'))
+
+
+
+

ps/trunk/source/tools/rlclient/python/zero_ad/environment.py

@@ -25 +25 @@
         self.current_state = GameState(json.loads(state_json), self)
         return self.current_state
+
+
+
 
     def get_template(self, name):