require(‘nn‘)

local function createQNetwork()
    local mlp = nn.Sequential()
    mlp:add(nn.Reshape(10))
    mlp:add(nn.Linear(10, 32))
    mlp:add(nn.Sigmoid())
    mlp:add(nn.Linear(32, 1))
    return mlp
end

local function qinput(obs_table)
    local obs = torch.Tensor(#obs_table):fill(0)
    for k = 1, #obs_table do
        obs[k] = obs_table[k]
    end
    obs = obs:view(#obs_table, -1) 
    return obs
end

qnn = createQNetwork()
for i = 1, 30 do
obs_table = {}
if i <10 then
  for j = 1, 10 do 
      obs_table[j] = 2
  end
else
  for j = 1, 10 do 
      obs_table[j] = 3
  end
end
obs = qinput(obs_table)
print(obs)
q = qnn:forward(obs)
print("q",q)
bs = torch.Tensor(1):fill(5)
cri = nn.MSECriterion()
qloss = cri:forward(q, bs)
dloss_dpredict = cri:backward(q, bs)
qnn:zeroGradParameters()
qgradInput = qnn:backward(obs, dloss_dpredict)
qnn:updateParameters(0.3)
end

簡單實例Torch7如何建立神經網絡