@@ -2,36 +2,45 @@ % http://www.trapexit.org/Erlang_and_Neural_Networks -module(ann). --export([perceptron/3, connect/2]). +-export([perceptron/4, connect/2]). -perceptron(Weights, Inputs, Sensitivities) -> +perceptron(Weights, Inputs, Sensitivities, Stale_inputs) -> Sigmoid = fun(X) -> 1/(1+math:exp(-X)) end, Sigmoid_deriv = fun(X) -> math:exp(-X)/(1+math:exp(-2*X)) end, receive {stimulate, Input} -> % add input to inputs + New_stale_inputs = delete_input(Stale_inputs, Input), New_inputs = replace_input(Inputs, Input), - {In_PID, In_Val} = Input, + {In_PID, In_Val, Expected_Val} = Input, ann_grapher ! {stimulate, In_PID, self(), In_Val}, - % calculate perceptron output - Output = feed_forward(Sigmoid, Weights, - convert_to_values(New_inputs)), - - % stimulate connected perceptrons - case Sensitivities of - [] -> - ann_grapher ! {output, self(), Output}, - io:format("Output (~w): ~w~n", [self(), Output]), - self() ! {learn, {self(), 1}}; - S -> lists:foreach(fun(Output_PID) -> - Output_PID ! {stimulate, {self(), Output}} - end, - convert_to_keys(S)) - end, - perceptron(Weights, New_inputs, Sensitivities); + + case New_stale_inputs of + % calculate perceptron output + [] -> Output = feed_forward(Sigmoid, Weights, + convert_to_input_values(New_inputs)), + + % stimulate connected perceptrons + case Sensitivities of + [] -> + ann_grapher ! {output, self(), Output}, + io:format("Output (~w): ~w~n", [self(), Output]), + case Expected_Val of + null -> ok; + E -> self() ! {learn, {self(), E}} + end; + S -> lists:foreach(fun(Output_PID) -> + Output_PID ! {stimulate, {self(), Output, Expected_Val}} + end, + convert_to_keys(S)) + end, + perceptron(Weights, New_inputs, Sensitivities, New_inputs); + + New_stale -> perceptron(Weights, New_inputs, Sensitivities, New_stale) + end; {learn, Backprop} -> @@ -41,16 +50,16 @@ perceptron(Weights, Inputs, Sensitivities) -> Learning_rate = 0.5, New_sensitivities = add_sensitivity(Sensitivities, Backprop), Output = feed_forward(Sigmoid, Weights, - convert_to_values(Inputs)), + convert_to_input_values(Inputs)), Deriv = feed_forward(Sigmoid_deriv, Weights, - convert_to_values(Inputs)), + convert_to_input_values(Inputs)), Sensitivity = calculate_sensitivities(Backprop, Inputs, New_sensitivities, Output, Deriv), Weight_adjustments = lists:map(fun(Input) -> Learning_rate * Sensitivity * Input end, - convert_to_values(Inputs)), + convert_to_input_values(Inputs)), New_weights = vector_map(fun(W, D) -> W+D end, Weights, Weight_adjustments), @@ -58,25 +67,26 @@ perceptron(Weights, Inputs, Sensitivities) -> Input_PID ! {learn, {self(), Sensitivity * Weight}} end, New_weights, - convert_to_keys(Inputs)), + convert_to_input_keys(Inputs)), - perceptron(New_weights, Inputs, New_sensitivities); + perceptron(New_weights, Inputs, New_sensitivities, Stale_inputs); {connect_to_output, Receiver_PID} -> Combined_sensitivities = - [{Receiver_PID, 0} | Sensitivities], - perceptron(Weights, Inputs, Combined_sensitivities); + [{Receiver_PID, 0.5} | Sensitivities], + perceptron(Weights, Inputs, Combined_sensitivities, Stale_inputs); {connect_to_input, Sender_PID} -> - Combined_input = [{Sender_PID, 0.5} | Inputs], - perceptron([0.5 | Weights], Combined_input, Sensitivities); + Combined_input = [{Sender_PID, 0.5, null} | Inputs], + New_stale_input = [{Sender_PID, 0.5, null} | Stale_inputs], + perceptron([0.5 | Weights], Combined_input, Sensitivities, New_stale_input); - {pass, Input_value} -> + {pass, Input_value, Expected_Output} -> lists:foreach(fun(Output_PID) -> - Output_PID ! {stimulate, {self(), Input_value}} + Output_PID ! {stimulate, {self(), Input_value, Expected_Output}} end, convert_to_keys(Sensitivities)), - perceptron(Weights, Inputs, Sensitivities) + perceptron(Weights, Inputs, Sensitivities, Stale_inputs) end. @@ -85,7 +95,7 @@ feed_forward(Func, Weights, Inputs) -> add_sensitivity([], _Backprop) -> []; add_sensitivity(Sensitivities, Backprop) -> - replace_input(Sensitivities, Backprop). + replace_sensitivity_input(Sensitivities, Backprop). calculate_sensitivities(_Backprop, [], _Sensitivities, _Output, _Deriv) -> @@ -102,10 +112,25 @@ connect(Sender, Receiver) -> Sender ! {connect_to_output, Receiver}, Receiver ! {connect_to_input, Sender}. -replace_input(Inputs, Input) -> +replace_sensitivity_input(Inputs, Input) -> {Input_PID, _} = Input, lists:keyreplace(Input_PID, 1, Inputs, Input). +replace_input(Inputs, Input) -> + {Input_PID, _, _} = Input, + lists:keyreplace(Input_PID, 1, Inputs, Input). + +delete_input(Inputs, Input) -> + {Input_PID, _, _} = Input, + lists:keydelete(Input_PID, 1, Inputs). + + +convert_to_input_values(Input_list) -> + lists:map(fun({_, Val, _}) -> Val end, Input_list). + +convert_to_input_keys(Input_list) -> + lists:map(fun({Key, _, _}) -> Key end, Input_list). + convert_to_values(Tuple_list) -> lists:map(fun({_, Val}) -> Val end, Tuple_list). ann.erl @@ -3,13 +3,13 @@ setup() -> ann_graph:start(), - N1_pid = spawn(ann, perceptron, [[],[],[]]), - N2_pid = spawn(ann, perceptron, [[],[],[]]), - N3_pid = spawn(ann, perceptron, [[],[],[]]), - N4_pid = spawn(ann, perceptron, [[],[],[]]), - N5_pid = spawn(ann, perceptron, [[],[],[]]), - N6_pid = spawn(ann, perceptron, [[],[],[]]), - N7_pid = spawn(ann, perceptron, [[],[],[]]), + N1_pid = spawn(ann, perceptron, [[],[],[],[]]), + N2_pid = spawn(ann, perceptron, [[],[],[],[]]), + N3_pid = spawn(ann, perceptron, [[],[],[],[]]), + N4_pid = spawn(ann, perceptron, [[],[],[],[]]), + N5_pid = spawn(ann, perceptron, [[],[],[],[]]), + N6_pid = spawn(ann, perceptron, [[],[],[],[]]), + N7_pid = spawn(ann, perceptron, [[],[],[],[]]), ann:connect(N1_pid, N2_pid), ann:connect(N1_pid, N3_pid), @@ -26,7 +26,7 @@ setup() -> ann:connect(N5_pid, N7_pid), ann:connect(N6_pid, N7_pid), - N1_pid ! {pass, 0.5}. + N1_pid ! {pass, 0.5, 0.7}. run() -> ann_graph:start(), ann_test.erl 226bebc11ee6dfbd5df8ff43b9d7ca516fd3919e Ben Hughes ben@pixelmachine.org http://github.com/schleyfox/erlang_ann/commit/98c108cdd47d9636799f0ee1a1a284c1e6d190bb 98c108cdd47d9636799f0ee1a1a284c1e6d190bb 2008-08-22T09:36:43-07:00 2008-08-22T09:36:43-07:00 Inputs can now be associated with the final expected value perceptrons now only fire when all inputs are new/fresh (this should help with overcorrection in training) 8866e26c4b17963de23e36108ba1fbf307029181 Ben Hughes ben@pixelmachine.org