Boilerplate/PoC of the Rete Algorithm implementated in Elixir
Rete is a complex stateful algorithm, this is an attempt of reproducing it with a functional immutable language such as Elixir/Erlang. This is a boilerplate that will let you implement your own version of Rete starting from a baseline.
Feel free to contact me via email if you have any question.
- Retex compiles the rules using
libgraph(A graph data structure library for Elixir projects) - The activation of nodes is done using a
Map.reduceover each node and activating them using theActivationprotocol. Each type of node implements its own protocol for the activation - When a node is activated, a map is updated adding its ID to the list of activated nodes in the Retex state
- A list of bindinds is passed along each iteration of the reduction on the graph and updated/checked depending on each implementation of the Activation protocol that the current node has
- Each production node should be visited only once
- Add tests for variables of the same name in different rules (it should work, right now it isnt)
- Improve performances and extend the facts to handle more complex cases such as relationships between WMEs
clone the project and run mix run benchmark/example.exs --no-mix-exs
The example adds 100_000 conditions and only one triggers a new inference with 6 working memory elements
You will see something like the following:
13:36:55.036 [info] Adding 100000 rules took 42 seconds, 508.98 milliseconds
13:36:55.432 [info] Network info %{num_edges: 400000, num_vertices: 299996, size_in_bytes: 278404360, type: :directed}
[agenda: [[{"$thing", :account_status, :silver}]]]
13:36:55.439 [info] Adding 6 working memories took 7.034 milliseconds and matched 1 rule
The code:
...
defp rule(n, type \\ :Thing) do
given = [
has_attribute(:Thing, :status, :==, "$a_#{n}"),
has_attribute(:Thing, :premium, :==, true)
]
action = [
{"$thing_#{n}", :account_status, "$a_#{n}"}
]
rule = create_rule(lhs: given, rhs: action)
end
def run() do
wme = Retex.Wme.new(:Account, :status, :silver)
wme_2 = Retex.Wme.new(:Account, :premium, true)
wme_3 = Retex.Wme.new(:Family, :size, 10)
wme_4 = Retex.Wme.new(:Account, :status, :silver)
wme_5 = Retex.Wme.new(:AccountC, :status, :silver)
wme_5 = Retex.Wme.new(:AccountD, :status, :silver)
number_of_rules = 100_000
require Logger
Logger.info("Adding #{number_of_rules} rules...")
# rules that will not match
result =
Timer.tc(fn ->
Enum.reduce(1..number_of_rules, Retex.new(), fn n, network ->
Retex.add_production(network, rule(n))
end)
end)
duration = result[:humanized_duration]
rules_100_000 = result[:reply]
Logger.info("Adding #{number_of_rules} rules took #{duration}")
# A rule that will match
given_matching = [
has_attribute(:Account, :status, :==, "$account"),
has_attribute(:Account, :premium, :==, true)
]
action_2 = [
{"$thing", :account_status, "$account"}
]
rule = create_rule(lhs: given_matching, rhs: action_2)
network = Retex.add_production(rules_100_000, rule)
Logger.info("Network info #{Graph.info(rules_100_000.graph) |> inspect()}")
# feed knowledge to the engine
result =
Timer.tc(fn ->
network
|> Retex.add_wme(wme)
|> Retex.add_wme(wme_2)
|> Retex.add_wme(wme_3)
|> Retex.add_wme(wme_4)
|> Retex.add_wme(wme_5)
end)
...
- So far it works very little and only in some cases, needs more testing and a better way to handle variable bindings
- This is just a WIP and a PoC, do not use it in production but feel free to contribute with ideas and PRs
- Use it as boilerplate and starting point to continue your own personal implementation (more facts, DSL, etc)
## For more on Rete algorithm