Skip to content

ML skills

The AEA ML (machine learning) skills demonstrate an interaction between two AEAs trading data.

There are two types of AEAs:

  • The ml_data_provider which sells training data.
  • The ml_model_trainer which trains a model

Discussion

The scope of the specific demo is to demonstrate how to create a simple AEA with integration of machine learning, and the usage of the AEA framework. The ml_data_provider AEA will provide some sample data and will deliver to the client upon payment. Once the client gets the data, it will train the model. The process can be found in the tasks.py file. This demo does not utilize a smart contract. As a result, we interact with a ledger only to complete a transaction.

Since the AEA framework enables us to use third-party libraries hosted on PyPI we can directly reference the external dependencies. The aea install command will install each dependency that the specific AEA needs and is listed in the skill's YAML file.

Communication

This diagram shows the communication between the two AEAs.

sequenceDiagram participant ml_model_trainer participant ml_data_provider participant Search participant Ledger activate ml_model_trainer activate ml_data_provider activate Search activate Ledger ml_data_provider->>Search: register_service ml_model_trainer->>Search: search Search-->>ml_model_trainer: list_of_agents ml_model_trainer->>ml_data_provider: call_for_terms ml_data_provider->>ml_model_trainer: terms ml_model_trainer->>Ledger: request_transaction ml_model_trainer->>ml_data_provider: accept (incl transaction_hash) ml_data_provider->>Ledger: check_transaction_status ml_data_provider->>ml_model_trainer: data loop train ml_model_trainer->>ml_model_trainer: tran_model end deactivate ml_model_trainer deactivate ml_data_provider deactivate Search deactivate Ledger

Preparation instructions

Dependencies

Follow the Preliminaries and Installation sections from the AEA quick start.

Demo instructions

Create data provider AEA

First, fetch the data provider AEA:

aea fetch fetchai/ml_data_provider:0.28.0
cd ml_data_provider
aea install
aea build

Alternatively, create from scratch.

The following steps create the data provider from scratch:

aea create ml_data_provider
cd ml_data_provider
aea add connection fetchai/p2p_libp2p:0.22.0
aea add connection fetchai/soef:0.23.0
aea add connection fetchai/ledger:0.18.0
aea add skill fetchai/ml_data_provider:0.25.0
aea config set --type dict agent.dependencies \
'{
  "aea-ledger-fetchai": {"version": "<2.0.0,>=1.0.0"}
}'
aea config set agent.default_connection fetchai/p2p_libp2p:0.22.0
aea config set --type dict agent.default_routing \
'{
  "fetchai/ledger_api:1.0.0": "fetchai/ledger:0.18.0",
  "fetchai/oef_search:1.0.0": "fetchai/soef:0.23.0"
}'
aea install
aea build

Create model trainer AEA

Then, fetch the model trainer AEA:

aea fetch fetchai/ml_model_trainer:0.29.0
cd ml_model_trainer
aea install
aea build

Alternatively, create from scratch.

The following steps create the model trainer from scratch:

aea create ml_model_trainer
cd ml_model_trainer
aea add connection fetchai/p2p_libp2p:0.22.0
aea add connection fetchai/soef:0.23.0
aea add connection fetchai/ledger:0.18.0
aea add skill fetchai/ml_train:0.26.0
aea config set --type dict agent.dependencies \
'{
  "aea-ledger-fetchai": {"version": "<2.0.0,>=1.0.0"}
}'
aea config set agent.default_connection fetchai/p2p_libp2p:0.22.0
aea config set --type dict agent.default_routing \
'{
  "fetchai/ledger_api:1.0.0": "fetchai/ledger:0.18.0",
  "fetchai/oef_search:1.0.0": "fetchai/soef:0.23.0"
}'
aea install
aea build

Add keys for the data provider AEA

First, create the private key for the data provider AEA based on the network you want to transact. To generate and add a private-public key pair for Fetch.ai AgentLand use:

aea generate-key fetchai
aea add-key fetchai fetchai_private_key.txt

Next, create a private key used to secure the AEA's communications:

aea generate-key fetchai fetchai_connection_private_key.txt
aea add-key fetchai fetchai_connection_private_key.txt --connection

Finally, certify the key for use by the connections that request that:

aea issue-certificates

Add keys and generate wealth for the model trainer AEA

The model trainer needs to have some wealth to purchase the data from the data provider.

First, create the private key for the model trainer AEA based on the network you want to transact. To generate and add a private-public key pair for Fetch.ai AgentLand use:

aea generate-key fetchai
aea add-key fetchai fetchai_private_key.txt

Then, create some wealth for your model trainer based on the network you want to transact with. On the Fetch.ai AgentLand network:

aea generate-wealth fetchai

Next, create a private key used to secure the AEA's communications:

aea generate-key fetchai fetchai_connection_private_key.txt
aea add-key fetchai fetchai_connection_private_key.txt --connection

Finally, certify the key for use by the connections that request that:

aea issue-certificates

Run both AEAs

Run both AEAs from their respective terminals.

First, run the data provider AEA:

aea run

Once you see a message of the form To join its network use multiaddr 'SOME_ADDRESS' take note of the address. (Alternatively, use aea get-multiaddress fetchai -c -i fetchai/p2p_libp2p:0.22.0 -u public_uri to retrieve the address.) This is the entry peer address for the local agent communication network created by the ML data provider.

Then, in the ML model trainer, run this command (replace SOME_ADDRESS with the correct value as described above):

aea config set --type dict vendor.fetchai.connections.p2p_libp2p.config \
'{
  "delegate_uri": "127.0.0.1:11001",
  "entry_peers": ["SOME_ADDRESS"],
  "local_uri": "127.0.0.1:9001",
  "log_file": "libp2p_node.log",
  "public_uri": "127.0.0.1:9001"
}'
This allows the model trainer to connect to the same local agent communication network as the data provider.

Then run the model trainer AEA:

aea run

You can see that the AEAs find each other, negotiate and eventually trade.

Cleaning up

When you're finished, delete your AEAs:

cd ..
aea delete ml_data_provider
aea delete ml_model_trainer