various work

main.lua

@@ -22,13 +22,19 @@ end
 
 -- configuration and globals.
 
-math.randomseed(10)
+--randomseed(11)
 
-local learning_rate = 1e-2
+local enable_overlay = false
-local deviation = 2e-2
+local enable_network = true
+--
+local epoch_trials = 24
+local learning_rate = 3.2e-3
+local deviation = 1e-1 / epoch_trials
+--
+local timer_loser = 1/3
+local cap_time = 400
 
 local epoch_i = 0
-local epoch_trials = 12
 local base_params
 local trial_i = 0
 local trial_noise = {}
@@ -38,17 +44,16 @@ local trials_remaining = 0
 local force_start = false
 local force_start_old = false
 
-local enable_overlay = false
-local enable_network = true
-
 local startsave = savestate.create(1)
 
 local poketime = false
+local max_time
 
 local sprite_input = {}
 local tile_input = {}
 
 local reward
+
 local powerup_old
 local status_old
 local coins_old
@@ -89,8 +94,10 @@ local ceil = math.ceil
 local min = math.min
 local max = math.max
 local exp = math.exp
+local log = math.log
 local sqrt = math.sqrt
 local random = math.random
+local randomseed = math.randomseed
 local insert = table.insert
 local remove = table.remove
 local unpack = table.unpack or unpack
@@ -177,12 +184,15 @@ local function make_network(input_size, buttons)
     nn_x = nn.Input(input_size)
     nn_y = nn_x
     nn_z = {}
+    if false then
         nn_y = nn_y:feed(nn.Dense(input_size))
+        nn_y = nn_y:feed(nn.Gelu())
+    else
+        nn_y = nn_y:feed(nn.Dense(114))
         nn_y = nn_y:feed(nn.Relu())
-    --nn_y = nn_y:feed(nn.Dense(floor(input_size / 16)))
+        nn_y = nn_y:feed(nn.Dense(57))
-    --nn_y = nn_y:feed(nn.Relu())
+        nn_y = nn_y:feed(nn.Relu())
-    --nn_y = nn_y:feed(nn.Dense(floor(input_size / 16)))
+    end
-    --nn_y = nn_y:feed(nn.Relu())
     for i = 1, buttons do
         nn_z[i] = nn_y
         nn_z[i] = nn_z[i]:feed(nn.Dense(2))
@@ -451,7 +461,7 @@ local function learn_from_epoch()
     print()
     print('rewards:', trial_rewards)
     normalize(trial_rewards)
-    print('normalized:', trial_rewards)
+    --print('normalized:', trial_rewards)
 
     local reward_mean, reward_dev = calc_mean_dev(trial_rewards)
 
@@ -465,7 +475,7 @@ local function learn_from_epoch()
     end
 
     local magnitude = learning_rate / deviation
-    print('stepping with magnitude', magnitude)
+    --print('stepping with magnitude', magnitude)
     -- throw the division from the averaging in there too.
     local altogether = magnitude / epoch_trials
     for i, v in ipairs(step) do
@@ -473,20 +483,23 @@ local function learn_from_epoch()
     end
 
     local step_mean, step_dev = calc_mean_dev(step)
-    print("step mean:", step_mean)
+    if step_dev < 1e-8 then
+        -- we didn't get anywhere. step in a random direction.
+        print("stepping randomly.")
+        local noise = trial_noise[1]
+        local devsqrt = sqrt(deviation)
+        for i, v in ipairs(step) do
+            step[i] = devsqrt * noise[i]
+        end
+
+        step_mean, step_dev = calc_mean_dev(step)
+    end
+    if abs(step_mean) > 1e-3 then print("step mean:", step_mean) end
     print("step stddev:", step_dev)
 
-    if step_dev > 1e-8 then
     for i, v in ipairs(base_params) do
         base_params[i] = v + step[i]
     end
-    else
-        -- we didn't get anywhere. step in a random direction.
-        local noise = trial_noise[1]
-        for i, v in ipairs(base_params) do
-            base_params[i] = v + magnitude * noise[i]
-        end
-    end
 
     network:distribute(base_params)
 
@@ -513,8 +526,9 @@ local function load_next_trial()
     print('loading trial', trial_i)
     local W = nn.copy(base_params)
     local noise = trial_noise[trial_i]
+    local devsqrt = sqrt(deviation)
     for i, v in ipairs(base_params) do
-        W[i] = v + deviation * noise[i]
+        W[i] = v + devsqrt * noise[i]
     end
     network:distribute(W)
 end
@@ -525,19 +539,11 @@ local function do_reset()
     if trial_i > 0 then trial_rewards[trial_i] = reward end
 
     if epoch_i == 0 or trial_i == epoch_trials then
-        if epoch_i > 0 then learn_from_epoch() else network:reset() end
+        if epoch_i > 0 then learn_from_epoch() end
         epoch_i = epoch_i + 1
         prepare_epoch()
     end
 
-    if once then
-        savestate.load(startsave)
-        print("end of trial reward:", reward)
-    else
-        savestate.save(startsave)
-    end
-    once = true
-
     -- bit of a hack:
     if get_state() == 'loading' then advance() end
     reward = 0
@@ -549,9 +555,20 @@ local function do_reset()
     -- unless you get a 1-up, in which case, please continue!
     W(0x75A, 0)
 
+    --max_time = min(log(epoch_i) * 10 + 100, cap_time)
+    max_time = min(8 * sqrt(15 * (epoch_i - 1)) + 100, cap_time)
+
+    if once then
+        savestate.load(startsave)
+        --print("end of trial reward:", reward)
+    else
+        savestate.save(startsave)
+    end
+    once = true
+
     emu.frameadvance() -- prevents emulator from quirking up.
 
-    print()
+    --print()
     load_next_trial()
 
     reset = false
@@ -564,7 +581,11 @@ local function init()
 
     emu.poweron()
     emu.unpause()
-    emu.speedmode("normal")
+    emu.speedmode("turbo")
+
+    network:load()
+    local res, err = pcall(network.load, network)
+    if res == false then print(err) end
 end
 
 init()
@@ -647,6 +668,8 @@ while true do
     local flagpole_bonus = R(0xE) == 4 and 1 or 0
     local reward_delta = screen_scroll_delta + status_delta * 256 + flagpole_bonus
 
+    -- TODO: add ingame score to reward.
+
     if not ingame_paused then reward = reward + reward_delta end
 
     --gui.text(4, 12, ("%02X"):format(#sprite_input), '#FFFFFF', '#0000003F')
@@ -655,7 +678,7 @@ while true do
     gui.text(112, 12, ("%+4i"):format(reward), '#FFFFFF', '#0000003F')
 
     if get_state() == 'dead' and state_old ~= 'dead' then
-        print("dead. lives remaining:", R(0x75A, 0))
+        --print("dead. lives remaining:", R(0x75A, 0))
         if R(0x75A, 0) == 0 then reset = true end
     end
     if get_state() == 'lose' then
@@ -664,17 +687,18 @@ while true do
     end
 
     -- lose a point for every frame paused.
-    if ingame_paused then reward = reward - 1 end
+    --if ingame_paused then reward = reward - 1 end
+    if ingame_paused then reward = reward - 402; reset = true end
 
     -- every few frames mario stands still, forcibly decrease the timer.
     -- this includes having the game paused.
     -- TODO: more robust. doesn't detect moonwalking against a wall.
     local timer = get_timer()
-    local timer_loser = 1/5
+    if ingame_paused or random() > 1 - timer_loser and R(0x1D) == 0 and R(0x57) == 0 then
-    if ingame_paused or math.random() > 1 - timer_loser and R(0x1D) == 0 and R(0x57) == 0 then
+        timer = timer - 1
-        timer = clamp(timer - 1, 0, 400)
-        set_timer(timer)
     end
+    timer = clamp(timer, 0, max_time)
+    set_timer(timer)
 
     -- if we've run out of time while the game is paused...
     -- that's cheating! unpause.

nn.lua

@@ -124,6 +124,7 @@ local Layer = Base:extend()
 local Model = Base:extend()
 local Input = Layer:extend()
 local Relu = Layer:extend()
+local Gelu = Layer:extend()
 local Dense = Layer:extend()
 local Softmax = Layer:extend()
 
@@ -241,6 +242,24 @@ function Relu:forward(X)
     return Y
 end
 
+function Gelu:init()
+    Layer.init(self, "Gelu")
+end
+
+function Gelu:forward(X)
+    assert(#X == self.size_in)
+    self.cache = self.cache or zeros(self.size_out)
+    local Y = self.cache
+
+    -- NOTE: approximate form of GELU exploiting similarities to sigmoid curve.
+    for i = 1, #X do
+        Y[i] = X[i] / (1 + exp(-1.704 * X[i]))
+    end
+
+    assert(#Y == self.size_out)
+    return Y
+end
+
 function Dense:init(dim)
     Layer.init(self, "Dense")
     assert(type(dim) == "number")
@@ -413,6 +432,7 @@ return {
     Model   = Model,
     Input   = Input,
     Relu    = Relu,
+    Gelu    = Gelu,
     Dense   = Dense,
     Softmax = Softmax,
 }