sharedRmsProp and async nature params

AsyncAgent.lua

@@ -2,15 +2,15 @@ local AsyncModel = require 'AsyncModel'
 local CircularQueue = require 'structures/CircularQueue'
 local classic = require 'classic'
 local optim = require 'optim'
-require 'modules/rmspropm' -- Add RMSProp with momentum
+require 'modules/sharedRmsProp'
 require 'classic.torch'
 
 local AsyncAgent = classic.class('AsyncAgent')
 
 local EPSILON_ENDS = { 0.01, 0.1, 0.5}
 local EPSILON_PROBS = { 0.4, 0.7, 1 }
 
-function AsyncAgent:_init(opt, policyNet, targetNet, theta, counters)
+function AsyncAgent:_init(opt, policyNet, targetNet, theta, counters, sharedG)
   log.info('creating AsyncAgent')
   local asyncModel = AsyncModel(opt)
   self.env, self.model = asyncModel:getEnvAndModel()
@@ -21,7 +21,8 @@ function AsyncAgent:_init(opt, policyNet, targetNet, theta, counters)
   self.optimiser = optim[opt.optimiser]
   self.optimParams = {
     learningRate = opt.eta,
-    momentum = opt.momentum
+    momentum = opt.momentum,
+    g = sharedG
   }
 
   local actionSpec = self.env:getActionSpec()

AsyncMaster.lua

@@ -1,3 +1,4 @@
+require 'socket'
 local AsyncModel = require 'AsyncModel'
 local AsyncAgent = require 'AsyncAgent'
 local class = require 'classic'
@@ -10,9 +11,10 @@ local TARGET_UPDATER = 1
 local VALIDATOR = 2
 
 local function checkNotNan(t)
-  local ok = t:ne(t):sum() == 0
+  local sum = t:sum()
+  local ok = sum == sum
   if not ok then
-    log.error('ERROR'.. t:sum())
+    log.error('ERROR'.. sum)
   end
   assert(ok)
 end
@@ -82,6 +84,7 @@ function AsyncMaster:_init(opt)
 
   self.theta = policyNet:getParameters()
   self.targetTheta = targetNet:getParameters()
+  local sharedG = self.theta:clone():zero()
 
   self.controlPool = threads.Threads(2)
   self.controlPool:specific(true)
@@ -93,7 +96,7 @@ function AsyncMaster:_init(opt)
   self.controlPool:addjob(VALIDATOR, torchSetup(opt))
   self.controlPool:addjob(VALIDATOR, function()
     local AsyncAgent = require 'AsyncAgent'
-    evalAgent = AsyncAgent(opt, policyNet, targetNet, theta, counters)
+    evalAgent = AsyncAgent(opt, policyNet, targetNet, theta, counters, sharedG)
   end)
 
   self.controlPool:synchronize()
@@ -113,7 +116,7 @@ function AsyncMaster:_init(opt)
       local mutex1 = threads1.Mutex(mutexId)
       mutex1:lock()
       local AsyncAgent = require 'AsyncAgent'
-      agent = AsyncAgent(opt, policyNet, targetNet, theta, counters)
+      agent = AsyncAgent(opt, policyNet, targetNet, theta, counters, sharedG)
       mutex1:unlock()
     end
   )
@@ -158,6 +161,8 @@ function AsyncMaster:start()
       local countSum = counters:sum()
       if countSum < 0 then return end
 
+      -- TODO report speed
+
       local countSince = countSum - lastUpdate
       if countSince > opt.tau then
         lastUpdate = countSum

async_main.lua

@@ -28,7 +28,7 @@ cmd:option('-doubleQ', 'true', 'Use Double Q-learning')
 -- Note from Georg Ostrovski: The advantage operators and Double DQN are not entirely orthogonal as the increased action gap seems to reduce the statistical bias that leads to value over-estimation in a similar way that Double DQN does
 cmd:option('-PALpha', 0.9, 'Persistent advantage learning parameter α (0 to disable)')
 -- Training options
-cmd:option('-optimiser', 'rmspropm', 'Training algorithm') -- RMSProp with momentum as found in "Generating Sequences With Recurrent Neural Networks"
+cmd:option('-optimiser', 'sharedRmsProp', 'Training algorithm')
 cmd:option('-eta', 0.0000625, 'Learning rate η') -- Prioritied experience replay learning rate (1/4 that of DQN; does not account for Duel as well)
 cmd:option('-momentum', 0.95, 'Gradient descent momentum')
 cmd:option('-batchSize', 5, 'Accumulate gradient x batchSize')
@@ -111,6 +111,7 @@ opt.Tensor = function(...)
   return torch.Tensor(...)
 end
 
+log.info(opt)
 
 local master = AsyncMaster(opt)
 

async_run.sh

@@ -33,7 +33,7 @@ if [ "$PAPER" == "demo" ]; then
   th async_main.lua -async $ASYNC -eta 0.00025 -doubleQ false -duel false -optimiser adam -steps 15000000 -tau 4 -epsilonSteps 10000 -valFreq 10000 -valSteps 6000 -PALpha 0 "$@"
 elif [ "$PAPER" == "nature" ]; then
   # Nature
-  th async_main.lua -async $ASYNC -game $GAME -duel false -tau 320000 -doubleQ false -PALpha 0 -eta 0.00025 -gradClip 0 "$@"
+    th async_main.lua -async $ASYNC -game $GAME -duel false -tau 40000 -optimiser sharedRmsProp -epsilonSteps 4000000 -doubleQ false -PALpha 0 -eta 0.0016 -gradClip 0 "$@"
 elif [ "$PAPER" == "doubleq" ]; then
   # Double-Q (tuned)
   th async_main.lua -async $ASYNC -game $GAME -duel false -PALpha 0 -eta 0.00025 -gradClip 0 "$@"

modules/sharedRmsProp.lua

@@ -0,0 +1,29 @@
+function optim.sharedRmsProp(opfunc, x, config, state)
+  -- Get state
+  local config = config or {}
+  local state = state or config
+  local lr = config.learningRate or 1e-2
+  local momentum = config.momentum or 0.95
+  local epsilon = config.epsilon or 0.01
+
+  -- Evaluate f(x) and df/dx
+  local fx, dfdx = opfunc(x)
+
+  -- Initialise storage
+  if not state.g then
+    state.g = torch.Tensor():typeAs(x):resizeAs(dfdx):zero()
+  end
+
+  if not state.tmp then
+    state.tmp = torch.Tensor():typeAs(x):resizeAs(dfdx)
+  end
+
+  state.g:mul(momentum):addcmul(1 - momentum, dfdx, dfdx)
+  state.tmp:copy(state.g):add(epsilon):sqrt()
+
+  -- Update x = x - lr x df/dx / tmp
+  x:addcdiv(-lr, dfdx, state.tmp)
+
+  -- Return x*, f(x) before optimisation
+  return x, {fx}
+end